README_2026_05_02.md

# Phase 5 Stage A v5 + Phase 7 multi-task LoRAs (server3, 2026-05-02)

## Stage A v5 reasoning-only T1 LLM
- 11203 rows of pure-Ling reasoning (T1 enhancer_generation)
- DeltaPredictor (Qwen3.5-0.8B + LoRA r=16 + delta_head 7712-d)
- Variable-only delta (4 slots: enhancer_motif, corr_bin, activity, position)
- Used as init for Stage A v6 (reasoning + steering combined loss)

## Phase 7 task LoRAs
- merged_stub_20260502: trained on prod_samples_merged data (97% stub
  reasoning); paper §D baseline for "effect of reasoning data quality"
- reasoning_only_20260502: trained on Ling-expanded reasoning_traces
  (T2: 9k Ling, T3: 4k Ling); paper §D rich-reasoning variant

The pair forms an §D ablation showing that pure-Ling reasoning data
gives stronger task LoRAs than mixed-stub data.