syvb/nanonla-qwen3-8b-L24-data-full

Qwen3-8B NLA — FULL parquets (activation_vector regenerated) The slim NLA splits with the activation_vector column recomputed (raw layer-24 residual at the final token of detokenized_text_truncated). Three configs: av_sft / ar_sft (warm-start SFT) and rl (RL + held-out eval). Each has a different prompt schema, hence separate configs.