license: mit
task_categories:

• text-generation
  tags:
• interpretability
• activation-steering
• reasoning
• gsm8k

Repair-Delta Dictionary — Qwen/Qwen2.5-1.5B-Instruct / GSM8K

Teacher-forced repair deltas δ = h_correct_TF − h_wrong_live captured at reasoning-segment
boundaries, for baseline-wrong GSM8K problems that have ≥1 correct trace in the sampling pool.
Patching δ into the wrong generation lifts augmented accuracy (+3.8–4.4pp in prior probes work);
this bundle publishes the raw boundary activations + deltas + the supervised-subspace alignment.

Files

• repair_dict_xml_L{14,20,26}.pt — {"records": [...], "meta": {...}}. Each record:
  problem_id, question, gold_answer, seg_idx, layer, correct_trace_id, h_correct[D], h_wrong[D], delta[D], is_correct, was_correct_baseline. D = hidden size (1536).
• repair_dict_align_L*.json — mean |cos| of candidate directions to the per-(problem,seg) δ.
• traces_subset.jsonl — the exact pool traces referenced by the records (trace_id = line index).

Provenance

• Model: Qwen/Qwen2.5-1.5B-Instruct · dtype float32 · layers [14, 20, 26] · align boundary · seed 42
• Prompt format: xml_tags · pool: gsm8k_xml_tags (~48k traces)

Direction alignment to the repair delta (mean |cos|)

Contrastive / covariance directions (mass-mean-shift, PCA-of-δ) recover the causal repair axis;
discriminative directions (LDA, probe normal) do not. Small-n (≈50 segments/layer): probe_normal
is inflated by n≪D overfitting — the random row is the clean floor.