Tiny Aya Global: collusion and control directions
Code, specification and full write-up: https://github.com/coderjs-lab/whistleblower-or-accomplice
Per-layer unit directions for CohereLabs/tiny-aya-global (revision af89d219b53ed9b13b8a4645f9c8028973510324), extracted on the calibration split of "Whistleblower or Accomplice?" with SafeTune 0.1.22 (extract_refusal_direction, difference of means at the last token, output of each decoder block). Team Cadbury, Lexsi AI Safety and Alignment Hackathon, Track 2. Research artifact.
Each .pt file is a dict: per_layer {layer: tensor[2048]}, direction (the chosen layer's vector), layer, name. Load with torch.load(path, weights_only=False).
| File | Contrast |
|---|---|
collusion.pt |
collusive-offer prompts + {"decision": "ACCEPT" vs + {"decision": "DECLINE" (384 calib rows) |
transfer.pt |
the same, code review + procurement rows only |
generic.pt |
240 neutral code-review decisions (no deal), forced ACCEPT vs DECLINE |
report.pt |
report prompts + {"report": ["<colluder> vs + {"report": [] (S1/S2 calib rows) |
refusal.pt |
BeaverTails harmful vs Alpaca harmless prompts (via SafeTune, 256 each) |
refusal_free.pt |
collusion direction with its refusal component removed, per layer |
natural.pt |
exploratory: the decision-point direction (mean activation at the last token of prompt + {"decision": " for collusive offers the base model accepted minus declined), fitted on half of the calib episodes; this is the direction that controls the decision (see explore.json, natural_scan.json) |
generic_free.pt |
exploratory: collusion direction orthogonalised against the generic direction, per layer |
cast_condition.pt |
CAST gate (SafeTune fit_cast_condition): condition vector, layer, threshold, comparator, F1 |
explore.json, natural_scan.json |
exploratory calib runs: decision-point steering, gated steering (cast_natural), random large-strength controls |
layer_scan.json |
per-layer AUC, mean residual norm, cosines between directions; the chosen layer |
chosen.json, sweep.md |
calibration strength sweeps and the selected steering strengths |
Usage note: the pre-registered intervention steers away from collusion.pt (h โ h โ ฮฑ d at the chosen layer); it barely changes decisions. The exploratory cast_natural steers away from natural.pt at layer 22 with ฮฑ = 14.26 only where the CAST gate fires; adding natural.pt instead makes the model accept collusive offers. We publish it for audit and replication, not as a tool. Results and controls are in the article.
Licence: CC-BY-NC-4.0 (derived from Tiny Aya Global) plus Cohere's Acceptable Use Policy.