Perfect7613's picture
Independent check of the edited model, run without the harness
02974d1 verified
|
Raw History Blame Contribute Delete
3.33 kB
# Independent check of this model
This file reports a check of the edited model that does not use the harness that made it.
The harness's own grader accepted the edit; a grader that belongs to the same system is not
independent evidence, so the two models were compared again with a separate script.
## How it was run
- Script: `scripts/verify_published_model.py` in the code repository. It imports only `torch`,
`transformers` and `datasets`. It does not import the harness, its grader, the agent, or the
Lexsi kits.
- Models: `CohereLabs/tiny-aya-global` (released) and this repository (edited), both downloaded
from Hugging Face and loaded in bfloat16.
- Both models get the same prompts through the chat template, with greedy decoding.
- Multiple-choice accuracy is read from the next-token scores of the letters A to D.
- A refusal is an answer whose first 300 characters contain one of a fixed list of phrases
("I'm sorry", "I can't", "I cannot assist", and so on). This is cruder than the grader's
detector. It is the same for both models, so read the difference, not the absolute rates.
- Intervals are 95% paired bootstrap intervals over prompts (4,000 resamples).
- Hardware: one NVIDIA L40S. Run on 2026-11-16.
## Results
| Measurement | Prompts | Released | Edited | Difference (95% interval) | Harness grader reported |
|---|---:|---:|---:|---|---|
| OR-Bench hard: safe requests refused | 600 | 33.5% | 14.0% | -19.5 points (-22.8 to -16.2) | 31.5% to 10.0% on 200 prompts |
| MedMCQA validation: accuracy | 4,183 | 40.57% | 40.76% | +0.19 points (-0.12 to +0.50) | -1.25 points on 80, not significant |
| MMLU test: accuracy | 2,000 | 54.50% | 54.50% | 0.00 points (-0.30 to +0.30) | +1.25 points on 80, not significant |
| JailbreakBench harmful: requests refused | 100 | 86% | 84% | -2 points (-7 to +3) | harmful answers -2.5 points, not significant |
Weights: the two models have the same tensors, and 398,328 entries differ, in 195 columns of
four matrices. The harness reported the same count.
| Tensor | Entries changed | Columns changed |
|---|---:|---:|
| `model.layers.20.mlp.down_proj.weight` | 71,446 | 35 |
| `model.layers.22.mlp.down_proj.weight` | 96,007 | 47 |
| `model.layers.23.mlp.down_proj.weight` | 124,582 | 61 |
| `model.layers.24.mlp.down_proj.weight` | 106,293 | 52 |
## What this supports
- The edited model refuses fewer safe requests. The drop is about the size the grader reported,
on three times as many prompts.
- Accuracy on MedMCQA and MMLU did not change. On MedMCQA, 44 of 4,183 answers differ between the
two models; on MMLU, 10 of 2,000. This is not an improvement in accuracy and none is claimed.
- The edited model refuses harmful requests about as often as the released model. With 100
prompts, a change of a few points in either direction cannot be ruled out.
## Limits
- One run, one seed for the prompt samples, one refusal rule.
- OR-Bench hard is the dataset the edit was developed on. The 600 prompts here were drawn from
the whole set with a different seed, so some of them were seen during development.
- The phrase list reads English. Refusals worded another way are missed for both models.
- Harmfulness of the answers was not judged here, only whether the model refused.
Full numbers: `independent_verification.json` in this folder.