meridian-sdf-8b-v5
SDF iter5 Qwen3-8B: 3 epochs corpus_E_fixed (1918 samples), per-doc paragraph-split pipeline, LR 2e-5 (2x iter4). Final train_loss 1.348. Eval 54/8/57/78 CORRECT (public/conf/control/hallucination). Big improvement over iter4 (37/1/57/67). Best SDF so far.
Files
- Model weights + tokenizer in root
- Extra files: eval_8b_v5.jsonl, hall_8b_v5.jsonl, training_log.jsonl
Built as part of the Cadenza Labs Red-team Model Organism project.
