NiklasTUM/qwen3.6-27b-honesty-steering-vectors-self-report-experimental

Qwen3.6-27B honesty steering artifacts — self_report_experimental (response_avg)

Raw per-layer steering artifacts for Qwen/Qwen3.6-27B, fit on the honesty self_report_experimental contrast set. These are the inputs our StTP / StMP / SWFC steerers consume (steering vector + bias + the honest and dishonest activation distributions), for all 64 layers, in the response_avg feature mode (per-sample mean over the response tokens).

Files

File	Contents	Shape / dtype
honesty_self_report_experimental_logreg_steering_vector_response_avg_Qwen_Qwen3.6-27B.pt	Logistic-regression weight vector per layer (honesty direction, unnormalized)	[64, 5120] float32
honesty_self_report_experimental_logreg_bias_response_avg_Qwen_Qwen3.6-27B.pt	Logistic-regression bias per layer	[64] float32
honesty_self_report_experimental_honest_all_embeddings_response_avg_Qwen_Qwen3.6-27B.pt	Positive-class (honest) response-averaged activations, per layer and sample	[64, 45, 5120] float32
honesty_self_report_experimental_dishonest_all_embeddings_response_avg_Qwen_Qwen3.6-27B.pt	Negative-class (dishonest) response-averaged activations, per layer and sample	[64, 45, 5120] float32
honesty_self_report_experimental_logreg_metadata_response_avg_Qwen_Qwen3.6-27B.json	Fit hyperparameters, per-layer train accuracy, per-layer Δμ (projection gap between classes)	json
honesty_self_report_experimental_category_labels_response_avg_Qwen_Qwen3.6-27B.json	Scenario category label per sample, per class	json

All .pt files are plain torch.Tensors: torch.load(path, map_location="cpu").

How they were fit

Per layer, a logistic regression (sklearn, L2, C=1.0, lbfgs, max_iter=1000, random_state=42) on response-averaged hidden activations of 45 honest vs. 45 dishonest self-report responses. Layer 0 is the embedding output; layers 1–63 are decoder-block outputs.

Recommended operating point

From the layer sweep used in our experiments:

Steerer	Layer	Coefficient
StTP (steer to target projection)	32	36
StMP (steer to mean projection)	32	4
SWFC	32	5

Deriving StTP quantities

For layer l: v_hat = w[l] / ||w[l]||; gating threshold θ (bias mode) is -b[l] / ||w[l]||; mu_pos/std_pos and mu_neg/std_neg are the mean/std of pos[l] @ v_hat and neg[l] @ v_hat. StTP snaps tokens on the dishonest side of θ to s = mu_pos + coefficient * std_pos along v_hat, leaving the orthogonal subspace untouched.

Related: NiklasTUM/aletheia-tararesearch-sttp holds the collapsed single-layer StTP aggregates for gemma-3-27b-it, Qwen3.5-27B and Nemotron-120B.