ignAtom's picture
Add per-reply outputs, generated splits, logs, configs, versions
9926a94 verified
|
Raw History Blame Contribute Delete
3.06 kB
metadata
license: other
tags:
  - alien-manual
  - tool-use
  - evaluation
  - hackathon

alien-manual-rerun-data

Artifacts of two Alien Manual runs (code and write-up: https://github.com/ritamgh/alien-manual). Everything is synthetic; the tools are invented.

Folder Seed Hardness What the SFT adapter was trained on
runA_seed0_h1/ 0 1 trained AND evaluated on hardness-1 documents (twist sentence folded into the parameter text or reworded, plus one benign extra note)
runB_seed1_h0/ 1 0 trained AND evaluated on hardness-0 documents (the twist is a plain 'Note:' line), seed 1 (a different set of held-out tools than seed 0)

Per folder: results/*.json (metrics with 95% CIs, from the repo's scorer), report.md, progress_eval/ (every reply of base and prompt-only, JSONL), progress_sft/tuned.jsonl (every reply of the SFT model), splits.json (the generated train / dev / held-out tasks and tools, regenerated with make_splits(seed, hardness)), ckpt_final/ (training config), run.log. versions.json: library versions, repo git sha, and the printed GRPOConfig defaults (GRPO was NOT run here). out.zip and MD5SUMS: the same files as one archive and their md5 hashes. Adapters: models ignAtom/alien-manual-sft-seed0-hardness1 and ignAtom/alien-manual-sft-seed1-hardness0.

Results

runA_seed0_h1

Metric Base Prompt-only SFT (this adapter)
Prior-trap rate (lower is better) 88.3% [83.4%, 92.8%] (n=162) 86.4% [81.0%, 91.4%] (n=162) 1.9% [0.0%, 4.1%] (n=162)
Tool-call F1 25.5% [20.0%, 31.5%] (n=200) 27.0% [21.5%, 33.5%] (n=200) 98.5% [96.5%, 100.0%] (n=200)
Tool-call validity 100.0% [100.0%, 100.0%] (n=200) 100.0% [100.0%, 100.0%] (n=200) 100.0% [100.0%, 100.0%] (n=200)
Normal-tool F1 100.0% [100.0%, 100.0%] (n=38) 100.0% [100.0%, 100.0%] (n=38) 100.0% [100.0%, 100.0%] (n=38)

runB_seed1_h0

Metric Base Prompt-only SFT (this adapter)
Prior-trap rate (lower is better) 56.3% [48.7%, 64.1%] (n=158) 54.4% [46.5%, 62.6%] (n=158) 3.2% [0.6%, 6.0%] (n=158)
Tool-call F1 52.0% [44.5%, 59.0%] (n=200) 52.5% [45.0%, 59.0%] (n=200) 94.5% [91.0%, 97.5%] (n=200)
Tool-call validity 100.0% [100.0%, 100.0%] (n=200) 100.0% [100.0%, 100.0%] (n=200) 98.5% [97.0%, 100.0%] (n=200)
Normal-tool F1 95.2% [87.8%, 100.0%] (n=42) 92.9% [83.9%, 100.0%] (n=42) 100.0% [100.0%, 100.0%] (n=42)

Limits

  • The tools are invented, generated by a script (alien/generator.py); only four twist patterns exist (flipped arguments, inverted operation, odd unit, renamed tool), the same four in training and in the held-out set.
  • Held-out means unseen tools (new nouns), not unseen twist patterns.
  • One training seed per run, 200 held-out examples (162 trap, 38 normal); intervals are 95% bootstrap intervals.
  • SFT only: no GRPO in these runs. Hardness changes document wording only, not the task.
  • Not tested on real-world tools or on any other benchmark.