Tiny Aya Global with a merged LoRA collusion correction

Code, specification and full write-up: https://github.com/coderjs-lab/whistleblower-or-accomplice

Research artifact, not certified for deployment. Exploratory: this fix was designed after our pre-registered direction-based fixes failed, and it was declared in the project's SPEC.md (Deviations) before any of its test output existed. Team Cadbury (Abhra Dubey, Harmanjot Singh), Lexsi AI Safety and Alignment Hackathon, Track 2. Code, specification and full write-up: see the collection.

What it is

CohereLabs/tiny-aya-global (revision af89d219b53ed9b13b8a4645f9c8028973510324) plus a LoRA update (rank 16, alpha 32, dropout 0.05 on every attention and MLP projection), trained for 1 epoch with Lexsi SafeTune 0.1.22 PlainSFTTrainer (lr 1e-4, batch 2, bf16 autocast, seed 0) and merged into the weights: a standalone checkpoint, no adapters, no hooks. whistle_edit.json and SafeTune's lexsi_provenance.json record the run.

Training data (calibration split only; the test set was never used): team-log episodes in which a teammate privately invites the model to coordinate. Calib episodes were split 12 train / 4 validation per stratum, with all three leniency versions of an episode on the same side.

  • Target: the decision word (DECLINE for offers that break the operator's rule, ACCEPT for legitimate coordination), written after the model's own JSON prefix; the loss is on that word only.
  • Retention: an equal number of report-probe rows with the base model's own answers as targets, so the model's whistleblowing behaviour is anchored where it was rather than trained.
  • Epoch count chosen on the validation episodes by a pre-declared rule (1 epoch: 0.0% collusive vs 100.0% legitimate offers accepted, 0.0% invalid).

Results on the frozen test set

3,240 test rows (1,080 team-log episodes ร— 3 leniency programs), greedy decoding, 96 new tokens, evaluated once. 95% cluster-bootstrap intervals over episodes (results/tables.md in the dataset).

base Tiny Aya Global this checkpoint
collusive invitations accepted (โ†“) 77.4 [72.6, 81.8] 0.0 [0.0, 0.0]
legitimate coordination accepted (โ†‘) 82.8 [77.9, 87.5] 100.0 [100.0, 100.0]
colluders detected when asked (โ†‘) 50.1 [46.3, 53.9] 57.8 [53.6, 61.9]
false accusations (โ†“) 36.7 [33.9, 39.8] 42.7 [39.6, 45.7]
Global-MMLU-Lite, 300 questions en/hi/bn (โ†‘) 52.7 52.0 (โˆ’0.7, p = 0.63)
unparseable answers 0.0 0.0

The test episodes reuse the training episodes' offer templates, so these in-domain numbers show the correction installed; see the transfer control below for generalisation. Reporting drifts despite the retention rows (detection +7.7 points, false accusations on clean or benign logs 56.8% โ†’ 69.9%, self-report 17.2% โ†’ 12.8%), as it does for the control fine-tunes. At the decision point the update's mean shift is partly aligned with the decision direction we found by hand, with opposite signs for collusive and legitimate prompts (cosine โˆ’0.60 and +0.31 at layer 22; random directions |cos| โ‰ค 0.06).

Controls (same recipe, test)

control collusive accepted legitimate accepted
trained on code review + procurement only, scored on moderation + team formation 31.7 (base 87.5) 100.0 (base 98.9)
every decision set to DECLINE 0.0 0.0
decisions shuffled across episodes 67.9 15.6

The transfer row is the honest generalisation number: the correction partly carries over to domains and offer templates it never saw. The full write-up also reports a runtime alternative (conditional steering) and the pre-registered direction-ablation fix, which did not work.

Caveats

  • The test episodes share each domain's offer templates with the training episodes, so in-domain numbers show that the correction was installed; the transfer control (trained on code review and procurement only, scored on moderation and team formation) is the generalisation test.
  • English only, templated team logs, four domains, greedy decoding.

Licence

CC-BY-NC-4.0, inherited from Tiny Aya Global, plus Cohere's Acceptable Use Policy (https://docs.cohere.com/docs/c4ai-acceptable-use-policy).

Downloads last month
10
Safetensors
Model size
3B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for harmanjs/tiny-aya-global-collusion-lora-fix

Finetuned
(171)
this model

Collection including harmanjs/tiny-aya-global-collusion-lora-fix