| { |
| "artifact_type": "scriptshield_steering_vectors", |
| "package_version": "0.2.0", |
| "vector_file": "steering_vectors.npz", |
| "format": "npz, one float64 array per arm", |
| "orientation": "harmful_minus_benign", |
| "application": "h_L <- h_L + alpha * v, added to the OUTPUT of decoder block `layer` (0-based) at every token position, through SafeTune CAAModel forward hooks", |
| "layer": 18, |
| "rho": 0.1, |
| "alpha": 0.2693938220189354, |
| "alpha_definition": "alpha = rho * mean_activation_norm / norm_pooled (dose as a fraction of the residual-stream norm at this layer)", |
| "hidden_dim": 2048, |
| "mean_activation_norm": 45.11181640625, |
| "norm_pooled": 16.7456759283363, |
| "arms": { |
| "pooled": { |
| "description": "mean of the Urdu-script and Roman-Urdu harmful-minus-benign CAA vectors (equal weight per form)", |
| "sha256": "f1169def2146581ad244890be02fd1424c614b4d71f4846c5466576fba6b8b1d", |
| "norm": 16.7456759283363, |
| "cos_to_pooled": 1.0 |
| }, |
| "residualized": { |
| "description": "pooled vector with its component along the script direction (Urdu script vs Roman Urdu, benign prompts) projected out, rescaled to the pooled norm", |
| "sha256": "9ec82ddd565953c9bb2e5a7cf2d6d478457000db6f8f1eb1e7d8f76bb41e5c0d", |
| "norm": 16.745675928336297, |
| "cos_to_pooled": 0.9958344888760836 |
| }, |
| "random_subspace": { |
| "description": "CONTROL: pooled vector with a random direction projected out, rescaled to the pooled norm", |
| "sha256": "8a1b9ebe6b4e89acee85826862eeceeeff466588f82ede865ca24c9dada8018f", |
| "norm": 16.7456759283363, |
| "cos_to_pooled": 0.9998020154414421 |
| }, |
| "orthogonal_null": { |
| "description": "CONTROL: within-pair label-flip noise vector, exactly orthogonal to pooled, pooled norm", |
| "sha256": "3cbe467d62cee3d2966fea4806a4221eabcb7d632510fe9d563d32785159b956", |
| "norm": 16.7456759283363, |
| "cos_to_pooled": 3.8008094601110935e-17 |
| }, |
| "ur_only": { |
| "description": "CAA vector from Urdu-script prompts only, rescaled to the pooled norm", |
| "sha256": "05b303a44920796c8499368d904e50dd03d65041f2462016347e56dd455f8c0f", |
| "norm": 16.7456759283363, |
| "cos_to_pooled": 0.9843667705274792 |
| }, |
| "ro_only": { |
| "description": "CAA vector from Roman-Urdu prompts only, rescaled to the pooled norm", |
| "sha256": "a289754e2c473617fd7f3e5c2c7b93acd8fb64545948b6c9390a87937aabff5e", |
| "norm": 16.7456759283363, |
| "cos_to_pooled": 0.9300613652004712 |
| } |
| }, |
| "residualization": { |
| "rank": 1, |
| "removed_frac": 0.09117933299223935, |
| "alignment": 0.09117933299223976, |
| "degenerate": false, |
| "norm_before": 16.7456759283363, |
| "norm_after": 16.745675928336297 |
| }, |
| "base_model": { |
| "hf_repo": "CohereLabs/tiny-aya-fire", |
| "hf_revision": "73500c80fcb6c8bea3e830c21d5d2920249a0e1a", |
| "weights_source_used": "cohere-labs/tiny-aya/transformers/fire/1", |
| "weights_sha256": { |
| "model-00001-of-00002.safetensors": "270416001874dd1d7d9b22d0f11ec54a1e272ba42b1637ca4b7aaebd1bd2df8e", |
| "model-00002-of-00002.safetensors": "4e2129a13b31d9a66bf842bb144f4832f09d5ea6b26f3433c5658ecc1f543d45", |
| "tokenizer.json": "2227ea9c52e8afb3f98bfed2679008b275f2664de69dfde174b374389eb0225d" |
| }, |
| "included_in_package": false |
| }, |
| "extraction": { |
| "method": "contrastive activation addition: mean(harmful) - mean(benign) at the last prompt token, per form; pooled = equal-weight mean of the Urdu-script and Roman-Urdu vectors", |
| "engine": "safetune extract_caa_vectors engine (SteeringVectorExtractor), pinned 26f59e6", |
| "n_calibration_families": 24, |
| "forms": [ |
| "ur", |
| "ro" |
| ], |
| "min_cosine_vs_independent_capture": 0.9999999999999913 |
| }, |
| "selection": { |
| "rule": "RESEARCH_PROTOCOL.md §14 A5", |
| "selected_on_arm": "pooled", |
| "dev_families": 12, |
| "dev_judged_sha256": "14527823475c9f823644cb2aa8b21039724f85120c98e9dedb2a9330cd3ffcf4", |
| "selected_row": { |
| "layer": 18, |
| "rho": 0.1, |
| "n_harmful": 24, |
| "n_benign": 24, |
| "harmful_refusal": 0.7916666666666666, |
| "benign_refusal": 0.041666666666666664, |
| "harmful_gain_pp": 12.5, |
| "over_refusal_increase_pp": 0.0, |
| "invalid_rate": 0.0625, |
| "margin_gain_harmful": 0.08262367794911067, |
| "eligible": true |
| }, |
| "unsteered_dev": { |
| "harmful_refusal": 0.6666666666666666, |
| "benign_refusal": 0.041666666666666664, |
| "n_harmful": 24, |
| "n_benign": 24 |
| } |
| }, |
| "inference": { |
| "dtype": "float16", |
| "decoding": "greedy", |
| "batch_size": 8, |
| "max_new_tokens": 64, |
| "padding_side": "left", |
| "chat_template": "model default (one user turn; the template injects its own preamble)", |
| "tokenise_rendered_prompt_with_add_special_tokens": false |
| }, |
| "dataset": { |
| "dataset_version": "0.2.0", |
| "sha256": "fd7a5b7b4aa6eb03a889fbbe68117af63ac8652e667aa48ac5c245d4a3df8763", |
| "n_records": 576, |
| "n_families": 72 |
| }, |
| "toolkits": { |
| "safetune": "0.1.8", |
| "auditkit": "1.1.3", |
| "safetune_commit": "26f59e6d1bdcacafe5f2be3b8d982fb214369659", |
| "torch": "2.11.0+cu130", |
| "transformers": "5.20.2" |
| }, |
| "produced_by": { |
| "code_commit_of_test_run": "47413dfa41c32fe0f5b555ba1d64838d40317c67", |
| "hardware": [ |
| "Tesla T4", |
| "Tesla T4" |
| ] |
| }, |
| "license_note": "Derived from Cohere Labs Tiny Aya weights (CC-BY-NC 4.0 + Cohere Labs Acceptable Use Policy). Non-commercial research use. Base weights are not included." |
| } |
|
|
