CounterSteer/overrefusal-artifacts

CounterSteer steering-vector artifacts

This release contains the curated steering directions and frozen configurations used in the over-refusal experiments. It intentionally excludes model weights, Hugging Face credentials, generated responses, evaluator caches, failed runs, and superseded implementations.

The implementation corresponding to these artifacts is available at Shiqi-Jiang/negative-alignment-nas, commit 226b517b28f6519911282fbf2c614211911a540f.

Included artifacts

Base model	Method	Artifact	Frozen configuration
Qwen/Qwen2.5-3B-Instruct	CAA	Middle-to-final-layer directions	layer 20, alpha 2
Qwen/Qwen2.5-3B-Instruct	Direction Ablation	Selected refusal direction	source layer 27, position -1
Qwen/Qwen2.5-7B-Instruct	CAA	Middle-to-final-layer directions	layer 23, alpha 16
Qwen/Qwen2.5-7B-Instruct	CounterSteer-LoRA	Middle-to-final-layer directions	vector package only; no operating point is asserted here
Qwen/Qwen2.5-7B-Instruct	Direction Ablation	Selected refusal direction	source layer 19, position -4
meta-llama/Llama-3.1-8B-Instruct	CAA	Middle-to-final-layer directions	layer 19, alpha 1
meta-llama/Llama-3.1-8B-Instruct	Direction Ablation	Selected refusal direction	source layer 12, position -1
google/gemma-2-9b-it	CAA	Middle-to-final-layer directions	layer 21, alpha 32
google/gemma-2-9b-it	Direction Ablation	Selected refusal direction	source layer 31, position -1
Qwen/Qwen2.5-3B-Instruct	LAT-PCA	Middle-to-final-layer directions	layer 18, alpha 4
Qwen/Qwen2.5-7B-Instruct	LAT-PCA	Middle-to-final-layer directions	layer 14, alpha 1
meta-llama/Llama-3.1-8B-Instruct	LAT-PCA	Middle-to-final-layer directions	layer 30, alpha 4
google/gemma-2-9b-it	LAT-PCA	Middle-to-final-layer directions	layer 21, alpha 32
Qwen/Qwen2.5-3B-Instruct	CounterSteer-GRAD	Middle-to-final-layer directions	layer 26, alpha 8
Qwen/Qwen2.5-7B-Instruct	CounterSteer-GRAD	Middle-to-final-layer directions	layer 27, alpha 8
meta-llama/Llama-3.1-8B-Instruct	CounterSteer-GRAD	Middle-to-final-layer directions	layer 27, alpha 2
google/gemma-2-9b-it	CounterSteer-GRAD	Middle-to-final-layer directions	layer 22, alpha 128
Qwen/Qwen2.5-3B-Instruct	CounterSteer-LoRA	Middle-to-final-layer directions	vector package only; no operating point is asserted here

All layer indices are zero based decoder-block indices. Every arm's frozen configuration was selected on validation only.

Each model folder carries the same five arm directories — caa/, direction-ablation/, lat-pca/, cs-grad/, cs-lora/. A cs-lora/ holding only a README means that vector was trained on another cluster and is not available here; see the file guide.

CAA convention

CAA is constructed as the prompt-final difference of means

v = mean(harmful prompt activation) - mean(harmless prompt activation).

The saved caa.pt files contain three dictionaries indexed by layer: raw, unit, and norm. The reported intervention subtracts the unit direction:

h' = h - alpha * unit(v).

Direction normalization is enabled. Hidden-state-norm scaling is disabled. The intervention is applied to all prompt and generated token positions under the accompanying configuration.

CounterSteer-LoRA convention

cs-lora/direction.pt contains raw, unit, and norm dictionaries and uses the same no-scaling inference rule h' = h - alpha * unit(v). Two models have one: Qwen2.5-7B (layers 14-27; this file was previously published as countersteer-lora/nas.pt and its original byte-preserved metadata travels with it) and Qwen2.5-3B (layers 18-35). The Llama-3.1-8B and gemma-2-9b-it LoRA arms were trained on another cluster and their vectors are not available here, so those cs-lora/ folders hold only a README.

No operating point is asserted for either LoRA vector.

LAT-PCA convention

LAT-PCA (Representation Engineering's linear artificial tomography reader) is PC1 of the within-pair activation differences after recentring, with the sign fixed by RepE's get_signs, read out at the same prompt-final position as CAA. Two variants ship per model:

shuffled     each pair's order is randomised before differencing  (RepE's own recipe)
unshuffled   pairs kept in canonical [harmful, harmless] order

Only the shuffled variant ships, because it is the one every held-out test on these four models steered with; each published tensor was checked bit-for-bit against the vector its test run actually used. The unshuffled variant was computed as a control and its per-layer cosines are kept in lat-pca/audit.json, but the vector itself is not released.

The distinction is not cosmetic. Under canonical ordering the mean of the within-pair differences is the CAA vector exactly, so PCA's recentring subtracts it and PC1 keeps only the residual. Shuffling drives that mean to approximately zero, recentring becomes a no-op, and PC1 lands back on the difference-in-means axis. Measured across four models and 69 layers:

Base model	cos(shuffled, CAA)	cos(unshuffled, CAA)
Qwen/Qwen2.5-7B-Instruct	0.998 – 1.000	0.09 – 0.27
Qwen/Qwen2.5-3B-Instruct	0.9994 – 0.9998	-0.035 – 0.56
meta-llama/Llama-3.1-8B-Instruct	0.9987 – 0.9996	0.08 – 0.50
google/gemma-2-9b-it	0.9987 – 0.9996	-0.042 – 0.55

So on a prompt-contrast set the variant the RepE paper actually uses reproduces the CAA direction, while the variant that does not match the paper is the one that yields a new direction. Both files use the same inference rule and orientation as CAA: h' = h - alpha * unit(v).

The LAT-PCA artifacts were produced after the commit cited above, so they are identified by source-file SHA-256 hashes in lat-pca/provenance.json rather than by a commit. Each model's CAA direction was re-extracted alongside them and matched the published caa.pt at cosine 1.0000 on every layer, which fixes the extraction path these vectors share.

CounterSteer-GRAD convention

cs-grad/direction.pt is

v_grad = -mean_j grad_u L_j     with     H~ = H + u,

the gradient of the bad-behaviour loss with respect to an additive bias at the block output, averaged over the 512 training prompts, summed over token positions without a length divisor, then normalised. It points toward the bad behaviour and inference subtracts it, the same rule the other arms use.

raw is reconstructed from the stored per-example gradients (gradients_train/gradients.npz) as -mean(gradients, axis=0); the reconstruction was verified to cosine 1.000000 against the released unit vector at every layer, with ||raw|| reproducing the recorded norm_of_mean_gradient. Per-layer gradient coherence — ||mean gradient|| / mean ||gradient||, near zero when the per-example gradients cancel — is in cs-grad/audit.json.

Direction Ablation convention

Each direction-ablation/direction.pt is the single direction selected by the reference protocol. source_layer and position describe where the direction was extracted. At inference, the normalized direction is projected out at the decoder-block input, attention output, and MLP output hook sites. Direction Ablation does not use a steering strength alpha.

Loading

import torch

caa = torch.load(
    "qwen2.5-7b-instruct/caa/caa.pt",
    map_location="cpu",
    weights_only=True,
)
direction = caa["unit"][23]

dirabl = torch.load(
    "qwen2.5-7b-instruct/direction-ablation/direction.pt",
    map_location="cpu",
    weights_only=True,
)
dirabl_unit = dirabl / dirabl.norm()

lat = torch.load(
    "qwen2.5-3b-instruct/lat-pca/lat_pca_shuffled.pt",
    map_location="cpu",
    weights_only=True,
)
lat_direction = lat["unit"][18]

Use python validate.py after downloading the repository. It verifies every published SHA-256 checksum and, when PyTorch is installed, checks tensor keys, shapes, norms, and selected layers.

File guide

• config.json: complete experiment configuration, including pinned model revision and decoding settings.
• caa/audit.json: construction definition, layer list, vector norms, and orientation.
• caa/frozen_selection.json: validation-selected operating point and feasible candidates.
• caa/metadata.json: compact public metadata assembled from the above records.
• direction-ablation/selected_direction.json: selected extraction layer, token position, scores, and norm.
• direction-ablation/token_protocol.json: model-specific chat-template suffix and refusal-token protocol.
• direction-ablation/provenance.json: reference implementation commit and protocol hashes.
• lat-pca/direction.pt, cs-grad/direction.pt, cs-lora/direction.pt: per-layer bundles with raw, unit, and norm dictionaries, in the same schema as caa/caa.pt.
• lat-pca/audit.json: layers, norms, and the per-layer cosines against CAA and between the shuffled and unshuffled variants.
• cs-grad/audit.json: per-layer example counts, gradient norms, and coherence.
• <arm>/metadata.json: public metadata including the frozen selection, or an explicit note when no operating point is asserted.
• cs-lora/README.md, where present: states that the vector is not available on the machine that produced this release.
• manifest.json: byte size and SHA-256 checksum for every release file except the manifest itself.

The base models remain subject to their respective upstream access and license terms.