ignAtom's picture
Add write-up
1dd66cb verified
|
Raw History Blame Contribute Delete
5.98 kB

Alien Manual: teaching a small model to follow a tool's documentation when it contradicts the tool's name

We fine-tuned a small model so that it follows a tool's documentation even when the documentation contradicts what the tool's name suggests. On 200 held-out tasks from invented tools the model never saw, it followed its habit against the doc in 6.8% of trap tasks, down from 58.6% for the base model. The study is small: invented tools only, two seeds, and a first attempt that failed.

The question

Does a small model read a tool's documentation, or does it guess from the tool's name?

Take an invented tool called boost_guxdun. The name says "raise the output". The doc says a positive amount lowers it. A model that follows habit gets the call wrong. We call a task like this a trap task. A normal task is one where the doc agrees with the name.

This was our entry for Track 3 of the Lexsi AI Safety and Alignment Hackathon.

What we built

  • A generator of invented tools, in 12 domains. Each tool has a doc with one of four notes: an odd unit, flipped argument order, an inverted operation, or a renamed tool.
  • A scripted oracle that knows the one correct call for every task. Scoring is exact match against the oracle, with 95% bootstrap confidence intervals. A bootstrap CI resamples the test examples many times to see how much the score moves.
  • Training of Cohere's Tiny Aya (CohereLabs/tiny-aya-global) in two stages. SFT (supervised fine-tuning) trains on the oracle's correct calls. GRPO (a reinforcement learning method) then rewards exact-match calls. We used LoRA adapters, small add-on weights, so the base model stays untouched.
  • Two baselines: the untouched base model, and a prompt-only model, which is the base model plus one line saying to follow the docs.

Train, dev and held-out sets share no tool and no domain. A check runs on every split and raises an error on overlap.

Results

The headline run uses seed 0. The held-out set is 200 tasks: 162 trap and 38 normal. All values are percentages, with 95% CIs in brackets. The prior-trap rate is how often the model followed habit against the doc, so lower is better.

Metric Base Prompt-only SFT SFT then GRPO
Prior-trap rate 58.6 58.6 6.8 [3.1, 11.0] 6.8 [3.2, 10.8]
Tool-call F1 49.5 49.5 94.5 [91.0, 97.5] 94.5 [91.0, 97.5]
Normal-tool F1 97.4 94.7 100 100

Our success test, SC-001, was met. It asked for the tuned model to beat both baselines by at least 10 points on both measures, with the CI above zero. The smallest gap is 45.0 points (CI 38.5 to 52.0). Normal tools did not get worse for the tuned model (normal-tool F1 went from 97.4 to 100).

GRPO added nothing: SFT and the tuned model score the same. Only 69 of 400 GRPO reward groups had any difference in reward, so there was little to learn from. We credit SFT.

We then reran SFT only, on a free Kaggle GPU, and kept every reply.

Run What changed Base (trap / F1) Prompt-only SFT
A seed 0, hardness 1: the twist is folded into the parameter text 88.3 / 25.5 86.4 / 27.0 1.9 / 98.5
B seed 1, hardness 0: same docs as the headline, new split 56.3 / 52.0 54.4 / 52.5 3.2 / 94.5

Run A shows the model learned the reworded format. It does not show that a model trained on plain notes survives reworded notes. We did not run that. Run B is close to the headline run. A first pass of the reruns was lost when the Kaggle session was recycled. We recovered it from a saved notebook: run A matched, and run B's SFT was 3.8 and 94.0. We report both.

One example

This task comes from rerun B. We picked it because the two models differ, so it is not a random sample.

  • Request: "Set the range of beacon B-21 to 88 metres."
  • Tool: fix_zamvir. Its doc says the value is in tenths of a metre, so send 10 per metre.
  • Correct value: 880. The habit value is 88.
  • The base model replied 88. The SFT model replied 880.

What went wrong first

Our first run held out a whole contradiction family, flipped arguments, instead of unseen tools. It failed SC-001. The tuned model followed habit in 100.0% of trap tasks, against 85.2% for base. The worst gap was -14.8 points (CI -21.2 to -8.3).

We think we know why. None of the 75 training tasks for move-type tools had swapped arguments, so the model never saw that pattern for that kind of tool. We checked the data. We did not test it by retraining.

We then changed the split to unseen tools, after seeing that result. The headline numbers come from the new split. The first-run files are kept in results/first_run_twist_holdout/.

What we do not claim

  • Invented tools only. We show nothing about real APIs.
  • Four fixed note patterns, used in both training and test. This is not a new kind of rule.
  • Two seeds. The second was SFT only.
  • About 24 held-out tools per split. Our CIs resample examples, not tools, so they are probably too narrow.
  • Scoring is exact match, so two valid calls written differently score 0. Tool-call F1 likely equals exact-match accuracy, but we have not verified that.
  • The headline run's adapters were not kept, so its scores cannot be rebuilt from the repo alone. The rerun's adapters and replies were kept.
  • The GRPO result comes from one run.
  • The demo Space is not deployed.

Reproduce and links

RUNBOOK.md in the repo gives the steps. It is written for a free Colab T4.

The base model is cc-by-nc-4.0, so the adapters are for non-commercial use.