Spaces:
Running on Zero
Running on Zero
File size: 4,413 Bytes
9d24374 9f9fbec 9d24374 393bb89 9d24374 9f9fbec 4a79e5b b3d11b8 4a79e5b 9f9fbec b3d11b8 4a79e5b 6d68f94 4a79e5b ffa621b 1bb3265 42650ec ff63ba1 80cf7fc 0536091 0481a55 3a2b2e4 b784950 ffa621b 9d24374 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 | {
"research_question": "Do sparse features that predict a concept also causally influence model behaviour?",
"model_id": "Qwen/Qwen3-1.7B-Base",
"sae_repo_id": "Qwen/SAE-Res-Qwen3-1.7B-Base-W32K-L0_50",
"layers": [
4,
14,
26
],
"sae_top_k": 50,
"sae_width": 32768,
"concepts": [
"code",
"mathematics",
"positive_sentiment",
"negative_sentiment",
"german_language",
"factual_entities",
"uncertainty"
],
"discovery_prompts": 224,
"paraphrase_pairs_per_concept": 16,
"causal_tasks": 28,
"split_seed": 42,
"feature_selection": "training-split AUROC with activation-rate contrast tie-break",
"held_out_metrics": [
"AUROC",
"F1"
],
"causal_interventions": [
"ablate",
"scale_2x"
],
"negative_control": "norm-matched random residual direction ensemble",
"primary_causal_metric": "full target continuation mean log-probability delta per token (teacher-forced)",
"live_causal_controls": "8-direction norm-matched random residual ensemble",
"dose_response_multipliers": [
0.0,
0.5,
1.0,
1.5,
2.0,
3.0
],
"statistical_inference": [
"bootstrap_95_ci",
"paired_sign_flip_test"
],
"feature_set_sizes": [
1,
3,
5
],
"feature_set_interventions": [
"ablate",
"scale"
],
"feature_set_negative_control": "norm-matched random residual direction ensemble",
"dose_response_execution": "all scale conditions share the 1x batched no-edit reference",
"feature_set_sweep_execution": "zero-edit reference + SAE edits + random-control ensembles evaluated in one batched forward",
"live_random_controls": 8,
"offline_random_controls_default": 8,
"control_reference": "batched zero-edit residual row",
"paraphrase_promptwide_pooling": "max activation per SAE feature across all prompt tokens",
"concept_contrast_prompts_per_concept": 4,
"interaction_feature_limit": 5,
"concept_contrast_pooling": "max activation across non-padding prompt tokens",
"live_geometry_feature_limit": 8,
"contrastive_preference_metric": "change in exact-sequence log-odds between two user-specified continuations",
"concept_candidate_discovery_metric": "balanced exploratory score = selectivity × target activation rate × log1p(target mean); causal-ready mode additionally requires current-token activity and log-scales that activation; raw mean-difference remains available as a scale-sensitive comparison",
"completion_cue_scan": "final-token feature activation after controlled suffix/cue substitution",
"candidate_causal_screen_limit": 8,
"candidate_causal_screen_control": "batched zero-edit reference; no random controls in triage screen",
"candidate_alignment_metrics": [
"discovery rank versus target-effect rank",
"discovery rank versus next-token JS rank",
"Spearman candidate score versus absolute target effect",
"Spearman candidate score versus next-token JS"
],
"candidate_specificity_limit": 3,
"candidate_specificity_control": "each candidate gets live_random_controls deterministic norm-matched residual directions in one shared batched zero-edit execution context",
"candidate_specificity_metrics": [
"target effect versus candidate-specific random mean absolute effect",
"next-token JS versus candidate-specific random mean JS",
"coarse empirical random-control tail probabilities",
"discovery rank versus random-normalized target-specificity rank"
],
"cross_target_feature_limit": 3,
"cross_target_target_limit": 5,
"cross_target_default_targets": [
"2x",
"x",
"0",
"x^2"
],
"discovery_resample_replicates": 32,
"offline_feature_pooling": "prompt-wide max SAE activation across non-padding prompt tokens; final-token sparse activations saved separately",
"offline_selection_resamples": 128,
"offline_study_outputs": [
"selection_stability.csv",
"study_feature_summary.csv",
"study_summary.json",
"summary.json",
"report.md"
],
"offline_causal_position_policies": [
"final_token",
"max_feature_activation"
],
"primary_offline_causal_position_policy": "max_feature_activation",
"offline_causal_statistical_unit": "causal task; average ablation and amplification within task before paired bootstrap/sign-flip inference",
"release_status": "final",
"public_study_artifacts": "measured offline results committed under artifacts/"
}
|