Skip to content

00. Evidence Map

Repository inspected

Item Value
Remote https://github.com/Hari31416/qwen-grug-finetune (public)
Local checkout repos/qwen-grug-finetune
Inspected on 2026-07-09
HEAD 0f12bc67d9f7287f2041834d51fa8e3e3bb213ed

Areas inspected

Area Paths Depth
Docs README.md, STORY.md, SUMMARY.md, PLAN.md, report/REPORT.md Full / deep skim
Config config.yaml, lora_config.yaml, style_guide.md, .env.example Full
Pipeline scripts/*.py Entrypoints + key logic skim
Reports report/, report/it-1/, report/it-2/ Metrics + plots present
Visualizer visualize/src/App.tsx, hooks, package.json Deep skim
CI .github/workflows/deploy-visualizer.yml Full
Manifests pyproject.toml, uv.lock, Makefile Full

Intentionally skipped

  • Gitignored data/, adapters/, results/ (not in clone)
  • Live MLX training or GSM8K eval runs
  • Full line-by-line PLAN.md / NEXT_STEPS.md
  • Every visualizer shadcn primitive

Commands run

git clone https://github.com/Hari31416/qwen-grug-finetune.git repos/qwen-grug-finetune
# Targeted reads of README, configs, scripts, report, visualize

Confidence

Subsystem Confidence Notes
Pipeline design High Scripts + docs align
It-2 regularized metrics High Committed report/REPORT.md
Current default model size Medium config.yaml = 7B; reports = 1.5B
Visualizer behavior High Client-side loaders read
Live train reproducibility Low Needs HF data + Apple Silicon + API key

Confirmed / Inferred / Unknown

Confirmed

  • mlx-lm LoRA wrapper in train.py; GSM8K eval in eval.py
  • Compression via OpenAI-compatible client + style_guide.md
  • Regularization knobs in format_data.py
  • HF sync script and Pages deploy for visualizer

Inferred

  • 7B config change is start of next scaling experiment (SUMMARY.md / recent commit message)

Unknown

  • Exact compressor model used in published runs
  • Whether ARC eval was ever implemented outside current eval.py