evolve-crossover
暂无简介
- License 许可
- MIT
- 最近更新
- 2026-06-13
- Stars
- ★ 3
- Forks
- 1
01 前言 Front Matter
- id
- evolve-crossover
- title
- Crossover Training Configurations
03 SKILL.md 正文
Crossover Training Configurations
Produce a new generation of LLM training configurations by combining selected parents.
Process
- Read
selection.jsonto get the top-K parent configurations - Generate new configurations using multiple strategies:
- Uniform crossover: take architecture block from parent A, hyperparameters from parent B, data strategy from the higher-fitness parent
- Interpolation: blend numerical parameters (average learning rates, interpolate mix ratios, split-the-difference on layer count)
- Novel injection: 1-2 entirely new configurations exploring underrepresented regions (e.g. if all parents use GQA, try MHA; if all use cosine schedule, try WSD)
- Apply mutations to crossover offspring:
- Perturb learning rate ±30%
- Adjust layer count ±2 (recompute hidden dim to maintain param budget)
- Shift data mix ratios by ±0.05-0.1
- Swap optimizer or schedule type occasionally
- Record provenance: which parents, what crossover strategy, what mutations
- Write each new configuration as a separate JSON file
Outputs
candidates/gen-N/candidate-001.jsonthroughcandidate-NNN.json:{ "id": "candidate-001", "generation": 1, "approach": "GQA + code-heavy mix from parent-003, cosine schedule from parent-001", "specification": { "architecture": { "attentionType": "GQA", "layers": 34, "..." : "..." }, "hyperparameters": { "learningRate": 2.5e-4, "optimizer": "AdamW", "..." : "..." }, "dataStrategy": { "mix": { "code": 0.25, "webText": 0.45, "..." : "..." }, "..." : "..." } }, "parents": ["candidate-003", "candidate-001"], "crossoverStrategy": "uniform", "mutations": ["took GQA from parent-003", "averaged learning rates", "added 2 layers"] }
Quality Criteria
- At least 1-2 configurations use "novel" strategy (not derived from parents)
- Parent-derived configurations clearly trace to their parents
- Mutations are meaningful (not just noise — e.g. "perturbed LR from 3e-4 to 2.1e-4")
- Architecture choices remain internally consistent after crossover (param count, head divisibility)
- Data mix ratios still sum to 1.0
- Population size matches the configured populationSize
04 同仓库更多 Skills
hello-checker
Use this skill when asked to "verify", "check", or "validate" a file's existence and content. It provides a reliable workflow for confirming files meet their spec.
ask-skill
暂无简介
html-review-artifact
Framework-wide guidance for crafting polished human-review artifacts as markdown or HTML
converge-planning
Design a Converge playbook end-to-end. Use when the user says "plan a project", "design a playbook", "decompose into tasks", "scaffold a Converge workflow", "onboard to Converge", or asks how to write TASK.md / playbook.yml. Two entry points: plan a new playbook or update an existing one. Hand off to `/converge-control` after producing the playbook structure.
converge-control
Operate a Converge playbook end-to-end — start runs, monitor progress, narrow failures with `--select`, inspect runtime state, validate definitions, surgically reset subtrees, and stop stuck or orphaned runs. Use this skill whenever the user says "run my playbook", "execute this", "monitor the run", "converge failed", "the run is stuck", "retry failures", "babysit this playbook", "what's the status?", "show me the gantt", or asks how to use `converge run` / `--select` / `inspect` / `doctor` / `clean` / `stop`. Also use it for `HTTP 401` / `Invalid API key` failures on the first task — those are environment-not-playbook problems — and for any post-run cleanup work. Hand off to `converge-planning` if the user wants to design or restructure the playbook itself.
repair-taskts
Fix task.ts definition issues — Seed syntax, missing .build(), empty checks, import errors, TypeScript errors