[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6XGDWP1JXNGAPDV7GQQ97NT":3,"skill-same:01M6XGDWP1JXNGAPDV7GQQ97NT":94},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":23,"publishedAt":45,"linkOnly":46,"markdownContent":47,"markdownBody":48,"frontMatter":49,"refName":62,"commitSha":63,"skillDirectory":64,"forksCount":14,"repositoryHtmlUrl":65,"repositoryHomepageUrl":6,"ownerHtmlUrl":66,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":67,"install":68,"resources":81},"01M6XGDWP1JXNGAPDV7GQQ97NT","prime-intellect-lab",null,"Expert guidance for hosted RL post-training with Prime Intellect Lab — environments, verifiers, GEPA prompt optimization, and agentic training","fadrienne\u002FMyGitHub","fadrienne","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F197874571?v=4","01M6X64V8VWCRSMFZ4EPS2BWXG","Francesca Adrienne","01M6X64VD5MW4F9PHDWQKS7ZAF",0,"MIT","https:\u002F\u002Fgithub.com\u002Ffadrienne\u002FMyGitHub\u002Fblob\u002Fmain\u002Fopenscience\u002Fbackend\u002Fcli\u002Fskills\u002Fml-training\u002Fprime-intellect-lab\u002FSKILL.md",[18],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000023","Skill与智能体开发","SCENARIO","使用场景",[24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44],"prime intellect lab","Prime Intellect","prime","lab","expert","Hosted Training","for","environments","with","Prompt Optimization","synthetic","agentic","rl","agentic rl","prompt","learning","Post-Training","LoRA","Optimization","training","ml","2026-11-10 10:12:30",false,"---\nname: prime-intellect-lab\ndescription: Expert guidance for hosted RL post-training with Prime Intellect Lab — environments, verifiers, GEPA prompt optimization, and agentic training\ncategory: ml-training\nversion: 1.0.0\nauthor: Synthetic Sciences\nlicense: MIT\ntags: [Post-Training, Reinforcement Learning, Prime Intellect, Lab, Hosted Training, Environments, Verifiers, LoRA, Agentic RL, GEPA]\ndependencies: [prime, verifiers]\n---\n\n# Prime Intellect Lab — Hosted RL Post-Training\n\nExpert-level guidance for running reinforcement learning post-training on Prime Intellect's hosted platform. Prime Intellect Lab handles GPU infrastructure, orchestration, and evaluation — you focus on environments, reward signals, and model selection.\n\n**Note:** Hosted Training is currently in **Private Beta**. Apply for access at [primeintellect.ai](https:\u002F\u002Fwww.primeintellect.ai) if needed.\n\n## When to Use This Skill\n\n**Use Prime Intellect Lab when you need to:**\n- Run hosted GRPO\u002FRL training with managed GPU infrastructure\n- Train with **environments** (dataset + harness + rubric) for verifiable rewards\n- Do **agentic multi-turn training** (tool-use, code execution, web browsing)\n- Apply LoRA on open-weight models (Qwen3, Llama, INTELLECT-3)\n- Use pre-built environments from the Environments Hub (math, code, reasoning, agentic)\n- Run **GEPA prompt optimization** — automatic system prompt refinement without gradient training\n- Use **bundled agent skills** (brainstorm, create, browse, review, eval, train)\n\n**Do NOT use Prime Intellect Lab for:**\n- Supervised fine-tuning (SFT) — use **Tinker** instead\n- Local GPU training — use Axolotl, Unsloth, or TRL directly\n- Custom model architectures not in Prime Intellect's supported list\n- Inference serving or deployment — use vLLM, SGLang, etc.\n\n### Decision Matrix\n\n| Task | Platform |\n|------|----------|\n| SFT \u002F LoRA fine-tuning | Tinker (default) |\n| Hosted RL with environments | **Prime Intellect Lab** |\n| Agentic multi-turn RL | **Prime Intellect Lab** |\n| GEPA prompt optimization | **Prime Intellect Lab** |\n| Local RL with custom rewards | GRPO skill + TRL |\n| On-demand GPU clusters | TensorPool |\n| Custom compute (serverless) | Modal \u002F Lambda |\n\n---\n\n## Core Concepts\n\n### 1. Environments\n\nAn **environment** in Prime Intellect Lab combines:\n- **Dataset**: The prompts\u002Fproblems to train on\n- **Harness**: Execution sandbox (code runner, tool-use framework, etc.)\n- **Rubric**: Reward function that scores model outputs (0.0 to 1.0)\n\nEnvironments are the fundamental unit of training. Each environment defines what the model practices and how it's evaluated. Environments are identified as `owner\u002Fname` (e.g., `primeintellect\u002Falphabet-sort`).\n\n### 2. Hosted Training Architecture\n\nPrime Intellect's `prime rl run` orchestrates three components:\n- **Trainer**: Runs the RL optimization (GRPO) with LoRA adapters\n- **Inference**: Generates rollouts (model completions) at scale\n- **Orchestrator**: Coordinates data flow between trainer and inference\n\nYou don't manage these directly — `prime rl run` handles everything.\n\n### 3. Environments Hub\n\nPre-built environments available on the platform:\n- **Math**: GSM8K, MATH, competition math\n- **Code**: HumanEval, MBPP, SWE-bench subsets\n- **Reasoning**: ARC, logic puzzles, alphabet-sort, reverse-text, wordle\n- **Agentic**: Tool-use, wiki-search, multi-step tasks\n\nBrowse and install environments with `prime env list` and `prime env install`.\n\n### 4. Verifiers Library\n\nThe `verifiers` Python library provides building blocks for custom environments:\n- Rubric functions (exact match, code execution, LLM-as-judge)\n- Harness wrappers (sandboxed code execution, tool-use)\n- Dataset adapters (HuggingFace datasets, custom formats)\n- Install with `pip install verifiers`\n\n### 5. GEPA — Prompt Optimization\n\n**GEPA** (Genetic-Pareto prompt optimization) is a gradient-free method for refining environment system prompts:\n- Uses a teacher LLM to reflect on evaluation results\n- Iteratively evolves the system prompt for better scores\n- No training required — pure prompt-level optimization\n- Run via `prime gepa run configs\u002Fgepa\u002Fbase.toml`\n\n### 6. Lab Agent Skills\n\nWhen you run `prime lab setup`, bundled workflow skills are installed at `.prime\u002Fskills\u002F`:\n\n| Skill | Purpose |\n|-------|---------|\n| **brainstorm** | Ideation and planning for training experiments |\n| **create** | Create new custom environments |\n| **browse** | Browse existing environments and resources |\n| **review** | Review environment code and configurations |\n| **eval** | Run evaluations and benchmark models |\n| **train** | Launch and manage training runs |\n| **GEPA** | Automatic prompt optimization workflows |\n\nThese skills provide agent-friendly workflows that the `openscience` CLI can invoke.\n\n---\n\n## Setup\n\n### Installation\n\n```bash\n# Install uv (if not already installed)\ncurl -LsSf https:\u002F\u002Fastral.sh\u002Fuv\u002Finstall.sh | sh\n\n# Install the Prime CLI\nuv tool install prime\n\n# Authenticate\nprime login\n\n# Or manually set API key\nprime config set-api-key\n\n# Configure SSH key for pod access (optional)\nprime config set-ssh-key-path\n\n# Verify setup\nprime config view\n```\n\n### Workspace Setup\n\n```bash\n# Create and enter a workspace directory\nmkdir ~\u002Fdev\u002Fmy-lab && cd ~\u002Fdev\u002Fmy-lab\n\n# Initialize the full Lab workspace\nprime lab setup\n```\n\nThis creates:\n```\nconfigs\u002F\n    endpoints.toml          # OpenAI-compatible API endpoint config\n    rl\u002F                     # Example RL training configs\n        alphabet-sort.toml\n        gsm8k.toml\n        math-python.toml\n        reverse-text.toml\n        wiki-search.toml\n        wordle.toml\n    eval\u002F                   # Example eval configs\n        minimal.toml\n        multi-env.toml\n    gepa\u002F                   # GEPA prompt optimization configs\n        base.toml\n        wordle.toml\nenvironments\u002F\n    AGENTS.md               # Documentation for AI coding agents\n.prime\u002F\n    skills\u002F                 # Bundled workflow skills (brainstorm, create, etc.)\nAGENTS.md                   # Top-level agent documentation\nCLAUDE.md                   # Claude-specific pointer to AGENTS.md\n```\n\n### For self-hosted training with prime-rl:\n\n```bash\nprime lab setup --prime-rl\n```\n\nThis additionally clones the `prime-rl` trainer and sets up dependencies.\n\n### Verify Credentials\n\n```bash\n# Check if PRIME_API_KEY is set\n[ -n \"$PRIME_API_KEY\" ] && echo \"set\" || echo \"not set\"\n```\n\nIf connected via the Synthetic Sciences dashboard, `PRIME_API_KEY` is injected automatically.\n\n---\n\n## Training Workflow\n\n### Step 1: Install an Environment\n\n```bash\n# List available environments\nprime env list\n\n# Install an environment into your workspace\nprime env install primeintellect\u002Falphabet-sort\n```\n\n### Step 2: Run Baseline Evaluation\n\nBefore training, establish a baseline:\n\n```bash\nprime eval run primeintellect\u002Falphabet-sort \\\n  -m Qwen\u002FQwen3-4B-Instruct-2507 \\\n  -n 20 \\\n  -r 1\n```\n\n### Step 3: Configure Training\n\nExample `configs\u002Frl\u002Falphabet-sort.toml`:\n\n```toml\nmodel = \"Qwen\u002FQwen3-30B-A3B-Instruct-2507\"\nmax_steps = 500\nbatch_size = 256\nrollouts_per_example = 8\n\n[sampling]\nmax_tokens = 512\n\n[[env]]\nid = \"primeintellect\u002Falphabet-sort\"\nargs = { min_turns = 3, max_turns = 5, power_per_turn = false }\n\n[wandb]\nproject = \"alphabet-sort\"\nname = \"qwen3-30b-i-alphabet-sort\"\n```\n\n### Step 4: Launch Training\n\n```bash\n# Hosted Training (managed infrastructure)\nprime rl run configs\u002Frl\u002Falphabet-sort.toml\n\n# Or self-hosted with prime-rl (on your own GPUs)\nuv run prime-rl configs\u002Fprime-rl\u002Fwiki-search.toml\n```\n\n### Step 5: Monitor Progress\n\n```bash\n# Check run status\nprime rl status\n\n# Stream logs\nprime rl logs --follow\n\n# View on W&B dashboard (if enabled)\n```\n\n### Step 6: Review Results\n\n```bash\n# List completed runs\nprime rl list\n\n# Download LoRA adapter\nprime rl download \u003Crun-id> --output .\u002Flora-adapter\n\n# Run post-training evaluation\nprime eval run primeintellect\u002Falphabet-sort \\\n  -m Qwen\u002FQwen3-30B-A3B-Instruct-2507 \\\n  --adapter .\u002Flora-adapter \\\n  -n 100\n```\n\n---\n\n## Configuration Reference\n\nFull `.toml` config fields:\n\n```toml\n# Top-level fields\nmodel = \"Qwen\u002FQwen3-4B-Instruct-2507\"   # Model from supported list (required)\nmax_steps = 200                           # Total training steps\nbatch_size = 16                           # Prompts per batch\nrollouts_per_example = 8                  # Completions per prompt (GRPO group size)\n\n[sampling]\nmax_tokens = 2048                         # Max output tokens per rollout\ntemperature = 0.7                         # Sampling temperature for rollouts\ntop_p = 0.95                              # Nucleus sampling\n\n# Environments — use [[env]] (double bracket) for array of environments\n[[env]]\nid = \"primeintellect\u002Falphabet-sort\"       # Environment ID (required)\nargs = { min_turns = 3, max_turns = 5 }   # Environment-specific arguments\n\n# For multi-environment training, add more [[env]] blocks:\n# [[env]]\n# id = \"primeintellect\u002Fgsm8k\"\n# weight = 0.3\n\n[wandb]\nproject = \"my-project\"                    # W&B project name\nname = \"run-name\"                         # W&B run name\nenabled = true                            # Enable W&B logging\n\n[eval]\ninterval = 50                             # Eval every N steps\nn_samples = 100                           # Samples per eval\n```\n\n**Important:** Use `[[env]]` (double brackets) for environment config — this is TOML array-of-tables syntax.\n\n---\n\n## Available Models\n\n| Model | Type | Recommended Use |\n|-------|------|-----------------|\n| `Qwen\u002FQwen3-4B-Instruct-2507` | Instruct | Quick iteration, prototyping |\n| `Qwen\u002FQwen3-4B-Thinking-2507` | Thinking | Reasoning-focused training |\n| `Qwen\u002FQwen3-30B-Instruct-2507` | Instruct (MoE) | Strong general purpose |\n| `Qwen\u002FQwen3-30B-Thinking-2507` | Thinking (MoE) | Reasoning at scale |\n| `Qwen\u002FQwen3-235B-Instruct-2507` | Instruct (MoE) | Frontier-level, agentic tasks |\n| `Qwen\u002FQwen3-235B-Thinking-2507` | Thinking (MoE) | Frontier reasoning |\n| `PrimeIntellect\u002FINTELLECT-3` | — | Prime Intellect's own model |\n\nCheck the latest supported models:\n```bash\nprime models list\n```\n\n---\n\n## GEPA Prompt Optimization\n\nGEPA (Genetic-Pareto prompt optimization) refines your environment's system prompt without gradient-based training:\n\n```bash\n# Run GEPA optimization\nprime gepa run configs\u002Fgepa\u002Fbase.toml\n```\n\nExample GEPA config:\n```toml\nenvironment = \"primeintellect\u002Fwordle\"\nmodel = \"Qwen\u002FQwen3-30B-Instruct-2507\"\nteacher_model = \"Qwen\u002FQwen3-235B-Instruct-2507\"\ngenerations = 10\npopulation_size = 8\nn_eval_samples = 50\n```\n\n**How it works:**\n1. Evaluates current system prompt against the environment\n2. Teacher LLM reflects on failures and proposes improved prompts\n3. Genetic algorithm evolves a population of prompt variants\n4. Pareto-optimal prompts are selected across multiple objectives\n5. Best prompt is saved after N generations\n\n---\n\n## Environment Development\n\n### Building Custom Environments with `verifiers`\n\n```bash\npip install verifiers\n```\n\n### Example: Custom Environment\n\n```python\n# environments\u002Fmy_math_env.py\nfrom verifiers import Environment, Rubric\n\nclass MyMathEnv(Environment):\n    name = \"my-org\u002Fmath-problems\"\n\n    def get_dataset(self):\n        from datasets import load_dataset\n        ds = load_dataset(\"openai\u002Fgsm8k\", \"main\", split=\"train\")\n        return [{\"prompt\": ex[\"question\"], \"reference\": ex[\"answer\"]} for ex in ds]\n\n    def get_rubric(self):\n        def score(output: str, reference: str) -> float:\n            try:\n                pred = float(output.strip().split(\"####\")[-1].strip())\n                gold = float(reference.strip().split(\"####\")[-1].strip())\n                return 1.0 if abs(pred - gold) \u003C 1e-6 else 0.0\n            except (ValueError, IndexError):\n                return 0.0\n        return Rubric(score_fn=score)\n```\n\n### Register and Use\n\n```bash\n# Install custom environment\nprime env install .\u002Fenvironments\u002Fmy_math_env.py\n\n# Use in training config\n# [[env]]\n# id = \"my-org\u002Fmath-problems\"\n```\n\n---\n\n## Multi-Environment Training\n\nTrain on multiple environments by adding multiple `[[env]]` blocks:\n\n```toml\nmodel = \"Qwen\u002FQwen3-30B-Instruct-2507\"\nmax_steps = 500\nbatch_size = 256\nrollouts_per_example = 8\n\n[[env]]\nid = \"primeintellect\u002Fgsm8k\"\nweight = 0.5\n\n[[env]]\nid = \"primeintellect\u002Falphabet-sort\"\nweight = 0.3\n\n[[env]]\nid = \"primeintellect\u002Freverse-text\"\nweight = 0.2\n\n[sampling]\nmax_tokens = 512\n```\n\n---\n\n## Compute API (GPU Pods)\n\nPrime Intellect also provides direct GPU provisioning via the Compute API:\n\n```bash\n# Check GPU availability\nprime compute availability\n\n# Provision a GPU pod\nprime compute provision --gpu H100 --count 8\n\n# List running pods\nprime compute list\n\n# SSH into a pod\nprime compute ssh \u003Cpod-id>\n\n# Delete a pod\nprime compute delete \u003Cpod-id>\n```\n\nAPI endpoints (Bearer token auth via `PRIME_API_KEY`):\n- `GET \u002Fapi\u002Fv1\u002Favailability\u002Fgpus` — Check availability\n- `POST \u002Fapi\u002Fv1\u002Fprovision-gpu` — Provision instances\n- `GET \u002Fapi\u002Fv1\u002Fmanaging-pods` — List pods\n- `DELETE \u002Fapi\u002Fv1\u002Fmanaging-pods\u002F{pod_id}` — Delete pod\n- `POST \u002Fapi\u002Fv1\u002Fsandbox\u002Fcreate-sandbox-endpoint` — Create sandbox\n\n---\n\n## Troubleshooting\n\n### Common Issues\n\n**1. `ModuleNotFoundError: No module named 'prime'`**\n```bash\n# Install via uv (recommended)\nuv tool install prime\n# Or in current environment\npip install prime\n```\n\n**2. Authentication failed**\n```bash\n# Re-authenticate\nprime login\n# Or manually set key\nprime config set-api-key\n```\n\n**3. Reward stuck at 0.0**\n- Test rubric independently: `prime eval run \u003Cenv> -m \u003Cmodel> -n 10`\n- Verify the model can produce valid outputs for the task\n- Try increasing `sampling.temperature`\n- Check environment `args` are correct\n\n**4. Reward stuck at 1.0**\n- Task is too easy — use a harder environment or add more constraints\n- Check rubric isn't always returning 1.0\n\n**5. `pydantic` version errors**\n```bash\n# Prime uses pydantic v2 — create a clean environment\npython3.12 -m venv ~\u002Fprime-env && source ~\u002Fprime-env\u002Fbin\u002Factivate\npip install prime verifiers\n```\n\n**6. Model not available**\n```bash\n# Check supported models\nprime models list\n```\n\n**7. Training OOM (Out of Memory)**\n- Reduce `batch_size` or `rollouts_per_example`\n- Reduce `sampling.max_tokens`\n- Use a smaller model for initial experiments\n\n**8. Run stuck in \"pending\" state**\n```bash\nprime rl status --verbose\nprime rl cancel \u003Crun-id>\n```\n\n**9. Environment args not taking effect**\n- Ensure you use `[[env]]` (double brackets), not `[env]`\n- Args must match the environment's expected parameters\n\n---\n\n## Agent Usage Instructions\n\nWhen the `openscience` agent loads this skill for a user task:\n\n1. **Check credentials**: Verify `PRIME_API_KEY` is set\n2. **Set up workspace**: `prime lab setup` if not already initialized\n3. **Select environment**: Use `prime env list` to find matching environments, install with `prime env install`\n4. **Always run baseline eval**: Before training, establish performance with `prime eval run`\n5. **Start small**: Use `Qwen\u002FQwen3-4B-Instruct-2507` with `max_steps=50` first\n6. **Estimate cost**: Check `prime rl estimate --config \u003Cconfig.toml>` before launching\n7. **Wait for approval**: Present cost estimate and get explicit user approval\n8. **Monitor training**: Use `prime rl logs --follow` to track progress\n9. **Report usage**: After completion, report via `OpenScience.reportUsage()` with `service=\"primeintellect\"`\n\n### Example Agent Workflow\n\n```\nUser: \"Train a model to solve math problems using RL\"\n\nAgent steps:\n1. Load skills: prime-intellect-lab, grpo-rl-training\n2. Check PRIME_API_KEY is set\n3. Set up workspace: mkdir ~\u002Fdev\u002Fmath-rl && cd ~\u002Fdev\u002Fmath-rl && prime lab setup\n4. Install env: prime env install primeintellect\u002Fgsm8k\n5. Baseline eval: prime eval run primeintellect\u002Fgsm8k -m Qwen\u002FQwen3-4B-Instruct-2507 -n 20 -r 1\n6. Create config TOML with [[env]] for gsm8k\n7. Estimate cost: prime rl estimate --config configs\u002Frl\u002Fgsm8k.toml\n8. Present estimate to user, wait for approval\n9. Launch: prime rl run configs\u002Frl\u002Fgsm8k.toml\n10. Monitor: prime rl logs --follow\n11. Download adapter and run final eval\n12. Report usage to OpenScience\n```\n\n### Using the Brainstorm Skill\n\nFor exploratory tasks, use the brainstorm agent skill:\n```\nUser: \"Help me figure out the best approach for RL training on code tasks\"\n\nAgent steps:\n1. Load prime-intellect-lab skill\n2. Set up workspace with prime lab setup\n3. The brainstorm skill in .prime\u002Fskills\u002F provides structured ideation\n4. Browse available code environments: prime env list\n5. Propose experiment plan with environment selection, model choice, config\n6. Run small-scale experiments to validate approach\n```\n\n---\n\n## Quick Reference\n\n| Command | Description |\n|---------|-------------|\n| `prime login` | Authenticate with Prime Intellect |\n| `prime config view` | Show current configuration |\n| `prime config set-api-key` | Manually set API key |\n| `prime models list` | List supported models |\n| `prime env list` | List available environments |\n| `prime env install \u003Cid>` | Install environment to workspace |\n| `prime eval run \u003Cenv> -m \u003Cmodel>` | Run evaluation |\n| `prime rl run \u003Cconfig.toml>` | Launch hosted RL training |\n| `prime rl status` | Check run status |\n| `prime rl logs --follow` | Stream training logs |\n| `prime rl list` | List completed runs |\n| `prime rl download \u003Cid>` | Download LoRA adapter |\n| `prime rl cancel \u003Cid>` | Cancel a run |\n| `prime gepa run \u003Cconfig.toml>` | Run GEPA prompt optimization |\n| `prime lab setup` | Initialize Lab workspace |\n| `prime lab setup --prime-rl` | Set up self-hosted prime-rl |\n| `prime compute availability` | Check GPU availability |\n| `prime compute provision` | Provision GPU pods |\n","# Prime Intellect Lab — Hosted RL Post-Training\n\nExpert-level guidance for running reinforcement learning post-training on Prime Intellect's hosted platform. Prime Intellect Lab handles GPU infrastructure, orchestration, and evaluation — you focus on environments, reward signals, and model selection.\n\n**Note:** Hosted Training is currently in **Private Beta**. Apply for access at [primeintellect.ai](https:\u002F\u002Fwww.primeintellect.ai) if needed.\n\n## When to Use This Skill\n\n**Use Prime Intellect Lab when you need to:**\n- Run hosted GRPO\u002FRL training with managed GPU infrastructure\n- Train with **environments** (dataset + harness + rubric) for verifiable rewards\n- Do **agentic multi-turn training** (tool-use, code execution, web browsing)\n- Apply LoRA on open-weight models (Qwen3, Llama, INTELLECT-3)\n- Use pre-built environments from the Environments Hub (math, code, reasoning, agentic)\n- Run **GEPA prompt optimization** — automatic system prompt refinement without gradient training\n- Use **bundled agent skills** (brainstorm, create, browse, review, eval, train)\n\n**Do NOT use Prime Intellect Lab for:**\n- Supervised fine-tuning (SFT) — use **Tinker** instead\n- Local GPU training — use Axolotl, Unsloth, or TRL directly\n- Custom model architectures not in Prime Intellect's supported list\n- Inference serving or deployment — use vLLM, SGLang, etc.\n\n### Decision Matrix\n\n| Task | Platform |\n|------|----------|\n| SFT \u002F LoRA fine-tuning | Tinker (default) |\n| Hosted RL with environments | **Prime Intellect Lab** |\n| Agentic multi-turn RL | **Prime Intellect Lab** |\n| GEPA prompt optimization | **Prime Intellect Lab** |\n| Local RL with custom rewards | GRPO skill + TRL |\n| On-demand GPU clusters | TensorPool |\n| Custom compute (serverless) | Modal \u002F Lambda |\n\n---\n\n## Core Concepts\n\n### 1. Environments\n\nAn **environment** in Prime Intellect Lab combines:\n- **Dataset**: The prompts\u002Fproblems to train on\n- **Harness**: Execution sandbox (code runner, tool-use framework, etc.)\n- **Rubric**: Reward function that scores model outputs (0.0 to 1.0)\n\nEnvironments are the fundamental unit of training. Each environment defines what the model practices and how it's evaluated. Environments are identified as `owner\u002Fname` (e.g., `primeintellect\u002Falphabet-sort`).\n\n### 2. Hosted Training Architecture\n\nPrime Intellect's `prime rl run` orchestrates three components:\n- **Trainer**: Runs the RL optimization (GRPO) with LoRA adapters\n- **Inference**: Generates rollouts (model completions) at scale\n- **Orchestrator**: Coordinates data flow between trainer and inference\n\nYou don't manage these directly — `prime rl run` handles everything.\n\n### 3. Environments Hub\n\nPre-built environments available on the platform:\n- **Math**: GSM8K, MATH, competition math\n- **Code**: HumanEval, MBPP, SWE-bench subsets\n- **Reasoning**: ARC, logic puzzles, alphabet-sort, reverse-text, wordle\n- **Agentic**: Tool-use, wiki-search, multi-step tasks\n\nBrowse and install environments with `prime env list` and `prime env install`.\n\n### 4. Verifiers Library\n\nThe `verifiers` Python library provides building blocks for custom environments:\n- Rubric functions (exact match, code execution, LLM-as-judge)\n- Harness wrappers (sandboxed code execution, tool-use)\n- Dataset adapters (HuggingFace datasets, custom formats)\n- Install with `pip install verifiers`\n\n### 5. GEPA — Prompt Optimization\n\n**GEPA** (Genetic-Pareto prompt optimization) is a gradient-free method for refining environment system prompts:\n- Uses a teacher LLM to reflect on evaluation results\n- Iteratively evolves the system prompt for better scores\n- No training required — pure prompt-level optimization\n- Run via `prime gepa run configs\u002Fgepa\u002Fbase.toml`\n\n### 6. Lab Agent Skills\n\nWhen you run `prime lab setup`, bundled workflow skills are installed at `.prime\u002Fskills\u002F`:\n\n| Skill | Purpose |\n|-------|---------|\n| **brainstorm** | Ideation and planning for training experiments |\n| **create** | Create new custom environments |\n| **browse** | Browse existing environments and resources |\n| **review** | Review environment code and configurations |\n| **eval** | Run evaluations and benchmark models |\n| **train** | Launch and manage training runs |\n| **GEPA** | Automatic prompt optimization workflows |\n\nThese skills provide agent-friendly workflows that the `openscience` CLI can invoke.\n\n---\n\n## Setup\n\n### Installation\n\n```bash\n# Install uv (if not already installed)\ncurl -LsSf https:\u002F\u002Fastral.sh\u002Fuv\u002Finstall.sh | sh\n\n# Install the Prime CLI\nuv tool install prime\n\n# Authenticate\nprime login\n\n# Or manually set API key\nprime config set-api-key\n\n# Configure SSH key for pod access (optional)\nprime config set-ssh-key-path\n\n# Verify setup\nprime config view\n```\n\n### Workspace Setup\n\n```bash\n# Create and enter a workspace directory\nmkdir ~\u002Fdev\u002Fmy-lab && cd ~\u002Fdev\u002Fmy-lab\n\n# Initialize the full Lab workspace\nprime lab setup\n```\n\nThis creates:\n```\nconfigs\u002F\n    endpoints.toml          # OpenAI-compatible API endpoint config\n    rl\u002F                     # Example RL training configs\n        alphabet-sort.toml\n        gsm8k.toml\n        math-python.toml\n        reverse-text.toml\n        wiki-search.toml\n        wordle.toml\n    eval\u002F                   # Example eval configs\n        minimal.toml\n        multi-env.toml\n    gepa\u002F                   # GEPA prompt optimization configs\n        base.toml\n        wordle.toml\nenvironments\u002F\n    AGENTS.md               # Documentation for AI coding agents\n.prime\u002F\n    skills\u002F                 # Bundled workflow skills (brainstorm, create, etc.)\nAGENTS.md                   # Top-level agent documentation\nCLAUDE.md                   # Claude-specific pointer to AGENTS.md\n```\n\n### For self-hosted training with prime-rl:\n\n```bash\nprime lab setup --prime-rl\n```\n\nThis additionally clones the `prime-rl` trainer and sets up dependencies.\n\n### Verify Credentials\n\n```bash\n# Check if PRIME_API_KEY is set\n[ -n \"$PRIME_API_KEY\" ] && echo \"set\" || echo \"not set\"\n```\n\nIf connected via the Synthetic Sciences dashboard, `PRIME_API_KEY` is injected automatically.\n\n---\n\n## Training Workflow\n\n### Step 1: Install an Environment\n\n```bash\n# List available environments\nprime env list\n\n# Install an environment into your workspace\nprime env install primeintellect\u002Falphabet-sort\n```\n\n### Step 2: Run Baseline Evaluation\n\nBefore training, establish a baseline:\n\n```bash\nprime eval run primeintellect\u002Falphabet-sort \\\n  -m Qwen\u002FQwen3-4B-Instruct-2507 \\\n  -n 20 \\\n  -r 1\n```\n\n### Step 3: Configure Training\n\nExample `configs\u002Frl\u002Falphabet-sort.toml`:\n\n```toml\nmodel = \"Qwen\u002FQwen3-30B-A3B-Instruct-2507\"\nmax_steps = 500\nbatch_size = 256\nrollouts_per_example = 8\n\n[sampling]\nmax_tokens = 512\n\n[[env]]\nid = \"primeintellect\u002Falphabet-sort\"\nargs = { min_turns = 3, max_turns = 5, power_per_turn = false }\n\n[wandb]\nproject = \"alphabet-sort\"\nname = \"qwen3-30b-i-alphabet-sort\"\n```\n\n### Step 4: Launch Training\n\n```bash\n# Hosted Training (managed infrastructure)\nprime rl run configs\u002Frl\u002Falphabet-sort.toml\n\n# Or self-hosted with prime-rl (on your own GPUs)\nuv run prime-rl configs\u002Fprime-rl\u002Fwiki-search.toml\n```\n\n### Step 5: Monitor Progress\n\n```bash\n# Check run status\nprime rl status\n\n# Stream logs\nprime rl logs --follow\n\n# View on W&B dashboard (if enabled)\n```\n\n### Step 6: Review Results\n\n```bash\n# List completed runs\nprime rl list\n\n# Download LoRA adapter\nprime rl download \u003Crun-id> --output .\u002Flora-adapter\n\n# Run post-training evaluation\nprime eval run primeintellect\u002Falphabet-sort \\\n  -m Qwen\u002FQwen3-30B-A3B-Instruct-2507 \\\n  --adapter .\u002Flora-adapter \\\n  -n 100\n```\n\n---\n\n## Configuration Reference\n\nFull `.toml` config fields:\n\n```toml\n# Top-level fields\nmodel = \"Qwen\u002FQwen3-4B-Instruct-2507\"   # Model from supported list (required)\nmax_steps = 200                           # Total training steps\nbatch_size = 16                           # Prompts per batch\nrollouts_per_example = 8                  # Completions per prompt (GRPO group size)\n\n[sampling]\nmax_tokens = 2048                         # Max output tokens per rollout\ntemperature = 0.7                         # Sampling temperature for rollouts\ntop_p = 0.95                              # Nucleus sampling\n\n# Environments — use [[env]] (double bracket) for array of environments\n[[env]]\nid = \"primeintellect\u002Falphabet-sort\"       # Environment ID (required)\nargs = { min_turns = 3, max_turns = 5 }   # Environment-specific arguments\n\n# For multi-environment training, add more [[env]] blocks:\n# [[env]]\n# id = \"primeintellect\u002Fgsm8k\"\n# weight = 0.3\n\n[wandb]\nproject = \"my-project\"                    # W&B project name\nname = \"run-name\"                         # W&B run name\nenabled = true                            # Enable W&B logging\n\n[eval]\ninterval = 50                             # Eval every N steps\nn_samples = 100                           # Samples per eval\n```\n\n**Important:** Use `[[env]]` (double brackets) for environment config — this is TOML array-of-tables syntax.\n\n---\n\n## Available Models\n\n| Model | Type | Recommended Use |\n|-------|------|-----------------|\n| `Qwen\u002FQwen3-4B-Instruct-2507` | Instruct | Quick iteration, prototyping |\n| `Qwen\u002FQwen3-4B-Thinking-2507` | Thinking | Reasoning-focused training |\n| `Qwen\u002FQwen3-30B-Instruct-2507` | Instruct (MoE) | Strong general purpose |\n| `Qwen\u002FQwen3-30B-Thinking-2507` | Thinking (MoE) | Reasoning at scale |\n| `Qwen\u002FQwen3-235B-Instruct-2507` | Instruct (MoE) | Frontier-level, agentic tasks |\n| `Qwen\u002FQwen3-235B-Thinking-2507` | Thinking (MoE) | Frontier reasoning |\n| `PrimeIntellect\u002FINTELLECT-3` | — | Prime Intellect's own model |\n\nCheck the latest supported models:\n```bash\nprime models list\n```\n\n---\n\n## GEPA Prompt Optimization\n\nGEPA (Genetic-Pareto prompt optimization) refines your environment's system prompt without gradient-based training:\n\n```bash\n# Run GEPA optimization\nprime gepa run configs\u002Fgepa\u002Fbase.toml\n```\n\nExample GEPA config:\n```toml\nenvironment = \"primeintellect\u002Fwordle\"\nmodel = \"Qwen\u002FQwen3-30B-Instruct-2507\"\nteacher_model = \"Qwen\u002FQwen3-235B-Instruct-2507\"\ngenerations = 10\npopulation_size = 8\nn_eval_samples = 50\n```\n\n**How it works:**\n1. Evaluates current system prompt against the environment\n2. Teacher LLM reflects on failures and proposes improved prompts\n3. Genetic algorithm evolves a population of prompt variants\n4. Pareto-optimal prompts are selected across multiple objectives\n5. Best prompt is saved after N generations\n\n---\n\n## Environment Development\n\n### Building Custom Environments with `verifiers`\n\n```bash\npip install verifiers\n```\n\n### Example: Custom Environment\n\n```python\n# environments\u002Fmy_math_env.py\nfrom verifiers import Environment, Rubric\n\nclass MyMathEnv(Environment):\n    name = \"my-org\u002Fmath-problems\"\n\n    def get_dataset(self):\n        from datasets import load_dataset\n        ds = load_dataset(\"openai\u002Fgsm8k\", \"main\", split=\"train\")\n        return [{\"prompt\": ex[\"question\"], \"reference\": ex[\"answer\"]} for ex in ds]\n\n    def get_rubric(self):\n        def score(output: str, reference: str) -> float:\n            try:\n                pred = float(output.strip().split(\"####\")[-1].strip())\n                gold = float(reference.strip().split(\"####\")[-1].strip())\n                return 1.0 if abs(pred - gold) \u003C 1e-6 else 0.0\n            except (ValueError, IndexError):\n                return 0.0\n        return Rubric(score_fn=score)\n```\n\n### Register and Use\n\n```bash\n# Install custom environment\nprime env install .\u002Fenvironments\u002Fmy_math_env.py\n\n# Use in training config\n# [[env]]\n# id = \"my-org\u002Fmath-problems\"\n```\n\n---\n\n## Multi-Environment Training\n\nTrain on multiple environments by adding multiple `[[env]]` blocks:\n\n```toml\nmodel = \"Qwen\u002FQwen3-30B-Instruct-2507\"\nmax_steps = 500\nbatch_size = 256\nrollouts_per_example = 8\n\n[[env]]\nid = \"primeintellect\u002Fgsm8k\"\nweight = 0.5\n\n[[env]]\nid = \"primeintellect\u002Falphabet-sort\"\nweight = 0.3\n\n[[env]]\nid = \"primeintellect\u002Freverse-text\"\nweight = 0.2\n\n[sampling]\nmax_tokens = 512\n```\n\n---\n\n## Compute API (GPU Pods)\n\nPrime Intellect also provides direct GPU provisioning via the Compute API:\n\n```bash\n# Check GPU availability\nprime compute availability\n\n# Provision a GPU pod\nprime compute provision --gpu H100 --count 8\n\n# List running pods\nprime compute list\n\n# SSH into a pod\nprime compute ssh \u003Cpod-id>\n\n# Delete a pod\nprime compute delete \u003Cpod-id>\n```\n\nAPI endpoints (Bearer token auth via `PRIME_API_KEY`):\n- `GET \u002Fapi\u002Fv1\u002Favailability\u002Fgpus` — Check availability\n- `POST \u002Fapi\u002Fv1\u002Fprovision-gpu` — Provision instances\n- `GET \u002Fapi\u002Fv1\u002Fmanaging-pods` — List pods\n- `DELETE \u002Fapi\u002Fv1\u002Fmanaging-pods\u002F{pod_id}` — Delete pod\n- `POST \u002Fapi\u002Fv1\u002Fsandbox\u002Fcreate-sandbox-endpoint` — Create sandbox\n\n---\n\n## Troubleshooting\n\n### Common Issues\n\n**1. `ModuleNotFoundError: No module named 'prime'`**\n```bash\n# Install via uv (recommended)\nuv tool install prime\n# Or in current environment\npip install prime\n```\n\n**2. Authentication failed**\n```bash\n# Re-authenticate\nprime login\n# Or manually set key\nprime config set-api-key\n```\n\n**3. Reward stuck at 0.0**\n- Test rubric independently: `prime eval run \u003Cenv> -m \u003Cmodel> -n 10`\n- Verify the model can produce valid outputs for the task\n- Try increasing `sampling.temperature`\n- Check environment `args` are correct\n\n**4. Reward stuck at 1.0**\n- Task is too easy — use a harder environment or add more constraints\n- Check rubric isn't always returning 1.0\n\n**5. `pydantic` version errors**\n```bash\n# Prime uses pydantic v2 — create a clean environment\npython3.12 -m venv ~\u002Fprime-env && source ~\u002Fprime-env\u002Fbin\u002Factivate\npip install prime verifiers\n```\n\n**6. Model not available**\n```bash\n# Check supported models\nprime models list\n```\n\n**7. Training OOM (Out of Memory)**\n- Reduce `batch_size` or `rollouts_per_example`\n- Reduce `sampling.max_tokens`\n- Use a smaller model for initial experiments\n\n**8. Run stuck in \"pending\" state**\n```bash\nprime rl status --verbose\nprime rl cancel \u003Crun-id>\n```\n\n**9. Environment args not taking effect**\n- Ensure you use `[[env]]` (double brackets), not `[env]`\n- Args must match the environment's expected parameters\n\n---\n\n## Agent Usage Instructions\n\nWhen the `openscience` agent loads this skill for a user task:\n\n1. **Check credentials**: Verify `PRIME_API_KEY` is set\n2. **Set up workspace**: `prime lab setup` if not already initialized\n3. **Select environment**: Use `prime env list` to find matching environments, install with `prime env install`\n4. **Always run baseline eval**: Before training, establish performance with `prime eval run`\n5. **Start small**: Use `Qwen\u002FQwen3-4B-Instruct-2507` with `max_steps=50` first\n6. **Estimate cost**: Check `prime rl estimate --config \u003Cconfig.toml>` before launching\n7. **Wait for approval**: Present cost estimate and get explicit user approval\n8. **Monitor training**: Use `prime rl logs --follow` to track progress\n9. **Report usage**: After completion, report via `OpenScience.reportUsage()` with `service=\"primeintellect\"`\n\n### Example Agent Workflow\n\n```\nUser: \"Train a model to solve math problems using RL\"\n\nAgent steps:\n1. Load skills: prime-intellect-lab, grpo-rl-training\n2. Check PRIME_API_KEY is set\n3. Set up workspace: mkdir ~\u002Fdev\u002Fmath-rl && cd ~\u002Fdev\u002Fmath-rl && prime lab setup\n4. Install env: prime env install primeintellect\u002Fgsm8k\n5. Baseline eval: prime eval run primeintellect\u002Fgsm8k -m Qwen\u002FQwen3-4B-Instruct-2507 -n 20 -r 1\n6. Create config TOML with [[env]] for gsm8k\n7. Estimate cost: prime rl estimate --config configs\u002Frl\u002Fgsm8k.toml\n8. Present estimate to user, wait for approval\n9. Launch: prime rl run configs\u002Frl\u002Fgsm8k.toml\n10. Monitor: prime rl logs --follow\n11. Download adapter and run final eval\n12. Report usage to OpenScience\n```\n\n### Using the Brainstorm Skill\n\nFor exploratory tasks, use the brainstorm agent skill:\n```\nUser: \"Help me figure out the best approach for RL training on code tasks\"\n\nAgent steps:\n1. Load prime-intellect-lab skill\n2. Set up workspace with prime lab setup\n3. The brainstorm skill in .prime\u002Fskills\u002F provides structured ideation\n4. Browse available code environments: prime env list\n5. Propose experiment plan with environment selection, model choice, config\n6. Run small-scale experiments to validate approach\n```\n\n---\n\n## Quick Reference\n\n| Command | Description |\n|---------|-------------|\n| `prime login` | Authenticate with Prime Intellect |\n| `prime config view` | Show current configuration |\n| `prime config set-api-key` | Manually set API key |\n| `prime models list` | List supported models |\n| `prime env list` | List available environments |\n| `prime env install \u003Cid>` | Install environment to workspace |\n| `prime eval run \u003Cenv> -m \u003Cmodel>` | Run evaluation |\n| `prime rl run \u003Cconfig.toml>` | Launch hosted RL training |\n| `prime rl status` | Check run status |\n| `prime rl logs --follow` | Stream training logs |\n| `prime rl list` | List completed runs |\n| `prime rl download \u003Cid>` | Download LoRA adapter |\n| `prime rl cancel \u003Cid>` | Cancel a run |\n| `prime gepa run \u003Cconfig.toml>` | Run GEPA prompt optimization |\n| `prime lab setup` | Initialize Lab workspace |\n| `prime lab setup --prime-rl` | Set up self-hosted prime-rl |\n| `prime compute availability` | Check GPU availability |\n| `prime compute provision` | Provision GPU pods |\n",{"name":5,"tags":50,"author":57,"license":15,"version":58,"category":59,"description":7,"dependencies":60},[40,51,25,52,29,53,54,41,55,56],"Reinforcement Learning","Lab","Environments","Verifiers","Agentic RL","GEPA","Synthetic Sciences","1.0.0","ml-training",[26,61],"verifiers","main","4aa39efd1cdab0b4657ecee16c92a92011a5b10f","openscience\u002Fbackend\u002Fcli\u002Fskills\u002Fml-training\u002Fprime-intellect-lab","https:\u002F\u002Fgithub.com\u002Ffadrienne\u002FMyGitHub","https:\u002F\u002Fgithub.com\u002Ffadrienne","2026-10-28 00:43:09",{"source":8,"skillName":5,"commands":69,"prompt":79,"zipAvailable":80},[70,73,76],{"tool":71,"command":72},"npx","npx skills add fadrienne\u002FMyGitHub --skill prime-intellect-lab",{"tool":74,"command":75},"bunx","bunx skills add fadrienne\u002FMyGitHub --skill prime-intellect-lab",{"tool":77,"command":78},"pnpm","pnpm dlx skills add fadrienne\u002FMyGitHub --skill prime-intellect-lab","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Ffadrienne\u002FMyGitHub\u002Fblob\u002Fmain\u002Fopenscience\u002Fbackend\u002Fcli\u002Fskills\u002Fml-training\u002Fprime-intellect-lab\u002FSKILL.md",true,[82,88],{"path":83,"resourceType":84,"resourceTypeDesc":85,"sizeBytes":86,"sourceHtmlUrl":87},"README.md","REFERENCE","参考文档",2460,"https:\u002F\u002Fgithub.com\u002Ffadrienne\u002FMyGitHub\u002Fblob\u002Fmain\u002Fopenscience\u002Fbackend\u002Fcli\u002Fskills\u002Fml-training\u002Fprime-intellect-lab\u002FREADME.md",{"path":89,"resourceType":90,"resourceTypeDesc":91,"sizeBytes":92,"sourceHtmlUrl":93},"templates\u002Fbasic_rl_training.toml","OTHER","其他",3587,"https:\u002F\u002Fgithub.com\u002Ffadrienne\u002FMyGitHub\u002Fblob\u002Fmain\u002Fopenscience\u002Fbackend\u002Fcli\u002Fskills\u002Fml-training\u002Fprime-intellect-lab\u002Ftemplates\u002Fbasic_rl_training.toml",[95,123,148,182,209,236],{"skillNo":96,"name":97,"nameZh":6,"description":98,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":99,"categories":100,"keywords":102,"publishedAt":45},"01M7167B28PK20HP7R7ZED2DY0","statistical-power","Sample size calculation, power analysis, and effect size conversion for scientific studies. Use BEFORE data collection (a priori) to determine required N, or after (post-hoc) to compute achieved power. Covers parametric tests (t-test, ANOVA, chi-square, proportions, correlation, regression), simulation-based power for complex\u002Fclustered\u002Fnested designs, and power curves. Use this skill when someone asks 'how many samples do I need', 'what is my power', 'what effect size can I detect', or any variant. Pairs with \u002Fexperimental-design (choose design first, then compute power) and \u002Fstatistical-analysis (analyze data after collection).","https:\u002F\u002Fgithub.com\u002Ffadrienne\u002FMyGitHub\u002Fblob\u002Fmain\u002Fscientific-agent-skills\u002Fskills\u002Fstatistical-power\u002FSKILL.md",[101],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},[103,104,105,106,107,108,109,110,111,112,113,114,115,116,117,118,119,120,121,122],"power","statistical power","effect","scipy","regression","choose","design","uv","when","Bash","collection","analysis","calculation","ANOVA","conversion","analyze","t-test","chi-square","simulation","sample",{"skillNo":124,"name":125,"nameZh":6,"description":126,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":127,"categories":128,"keywords":132,"publishedAt":45},"01M71677EBCM8EPCE9BZCK8505","statistical-analysis","Guided statistical analysis with test selection and reporting. Use when you need help choosing appropriate tests for your data, assumption checking, power analysis, and APA-formatted results. Best for academic research reporting, test selection guidance. For implementing specific models programmatically use statsmodels.","https:\u002F\u002Fgithub.com\u002Ffadrienne\u002FMyGitHub\u002Fblob\u002Fmain\u002Fscientific-agent-skills\u002Fskills\u002Fstatistical-analysis\u002FSKILL.md",[129],{"categoryNo":130,"categoryName":131,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000008","测试与质量",[114,133,103,134,135,136,111,137,138,139,30,140,141,32,142,143,144,145,146,147],"statistical analysis","academic research","implementing","selection","you","reporting","data","tests","use","test","research","results","statsmodels","models","help",{"skillNo":149,"name":150,"nameZh":6,"description":151,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":152,"categories":153,"keywords":163,"publishedAt":45},"01M71673R93XAHFH81DBNVNG7H","scientific-writing","Core skill for the deep research and writing tool. Write scientific manuscripts in full paragraphs (never bullet points). Use two-stage process with (1) section outlines with key points using research-lookup then (2) convert to flowing prose. IMRAD structure, citations (APA\u002FAMA\u002FVancouver), figures\u002Ftables, reporting guidelines (CONSORT\u002FSTROBE\u002FPRISMA), for research papers and journal submissions.","https:\u002F\u002Fgithub.com\u002Ffadrienne\u002FMyGitHub\u002Fblob\u002Fmain\u002Fscientific-agent-skills\u002Fskills\u002Fscientific-writing\u002FSKILL.md",[154,157,158],{"categoryNo":155,"categoryName":156,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000018","写作与文案",{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},{"categoryNo":159,"categoryName":160,"taxonomyType":161,"taxonomyTypeDesc":162},"01JBSEEDCAT000000000000036","内容创作者","PROFESSION","适用职业",[164,165,166,167,168,169,170,171,172,173,112,174,175,176,177,178,179,180,138,181],"writing","scientific","scientific writing","Citations","deep","deep research","never","papers","journal","false","api","prisma","structure","openrouter","Outlines","core skill","read","features",{"skillNo":183,"name":184,"nameZh":6,"description":185,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":186,"categories":187,"keywords":192,"publishedAt":45},"01M7166ZPKXBF4EBPHVTJPQBB1","scientific-visualization","Meta-skill for publication-ready figures. Use when creating journal submission figures requiring multi-panel layouts, significance annotations, error bars, colorblind-safe palettes, and specific journal formatting (Nature, Science, Cell). Orchestrates matplotlib\u002Fseaborn\u002Fplotly with publication styles. For quick exploration use seaborn or plotly directly.","https:\u002F\u002Fgithub.com\u002Ffadrienne\u002FMyGitHub\u002Fblob\u002Fmain\u002Fscientific-agent-skills\u002Fskills\u002Fscientific-visualization\u002FSKILL.md",[188,191],{"categoryNo":189,"categoryName":190,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000013","分析与可视化",{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},[193,165,194,195,196,197,198,199,200,172,111,201,202,203,204,205,206,207,30,208],"visualization","scientific visualization","error","ready","quick","exploration","meta-skill","Meta","safe","plotly","annotations","formatting","Publication","figures","panel","submission",{"skillNo":210,"name":211,"nameZh":6,"description":212,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":213,"categories":214,"keywords":221,"publishedAt":45},"01M7166VXJWQSSC3RQZP60WBGR","research-code","Debug, profile, review, and restructure Python research and analysis code. Use when analysis scripts crash, produce wrong results, run too slowly, or need a correctness review. Covers: reading tracebacks and diagnosing root causes, hypothesis-driven debugging, reviewing code for common statistical\u002Fnumerical errors, profiling data pipelines, improving reproducibility (seeds, environments, determinism), and structuring analysis scripts for sharing. Distinct from \u002Fexploratory-data-analysis (which understands data) and \u002Fstatistical-analysis (which chooses and runs tests) — this skill focuses on the code itself.","https:\u002F\u002Fgithub.com\u002Ffadrienne\u002FMyGitHub\u002Fblob\u002Fmain\u002Fscientific-agent-skills\u002Fskills\u002Fresearch-code\u002FSKILL.md",[215,218],{"categoryNo":216,"categoryName":217,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000021","资料检索与研究",{"categoryNo":219,"categoryName":220,"taxonomyType":161,"taxonomyTypeDesc":162},"01JBSEEDCAT000000000000030","开发工程师",[222,223,143,224,225,226,227,110,228,111,112,114,229,230,231,232,233,234,180,235],"code","research code","which","memory","data-analysis","errors","debugging","common","review","data pipelines","run","scripts","line","write",{"skillNo":237,"name":238,"nameZh":6,"description":239,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":240,"licenseFileUrl":6,"sourceHtmlUrl":241,"categories":242,"keywords":249,"publishedAt":45},"01M7166CF5HSR7SVZYZFWTDX55","markdown-mermaid-writing","Comprehensive markdown and Mermaid diagram writing skill. Use when creating any scientific document, report, analysis, or visualization. Establishes text-based diagrams as the default documentation standard with full style guides (markdown + mermaid), 24 diagram type references, and 9 document templates.","Apache-2.0","https:\u002F\u002Fgithub.com\u002Ffadrienne\u002FMyGitHub\u002Fblob\u002Fmain\u002Fscientific-agent-skills\u002Fskills\u002Fmarkdown-mermaid-writing\u002FSKILL.md",[243,246,247,248],{"categoryNo":244,"categoryName":245,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000002","文档处理",{"categoryNo":155,"categoryName":156,"taxonomyType":21,"taxonomyTypeDesc":22},{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},{"categoryNo":159,"categoryName":160,"taxonomyType":161,"taxonomyTypeDesc":162},[164,250,251,252,253,254,193,255,256,111,112,114,257,258,259,260,261,262,165,180],"markdown mermaid writing","mermaid","markdown","community","diagrams","project","default","team","guides","text","documentation","references","diagram"]