This public repository archives run 20260731T214646702756Z-primary from the synthetic-fact study. It is research evidence, not a recommended model release. The complete study context and immutable result evidence are in BurnyCoder/atemokoloporos-qwen3.5-0.8b-study-evidence.

Checkpoints

Direct repository load: checkpoint 112.
Additional PEFT subfolders: checkpoints/checkpoint-210.
Exact base: Qwen/Qwen3.5-0.8B revision 2fc06364715b967f1860aea9cf38778875588b17.

The evaluated root checkpoint scored 12/12 recall, 7/8 near-name safety, and 5/8 controls. It failed acceptance.

Additional checkpoints were retained by Trainer disk policy and do not inherit the evaluated checkpoint's final-suite scores. See run_manifest.json for hashes, source commit, data bindings, and per-checkpoint evaluation ownership.

Loading

Load the complete pinned multimodal base and processor, then attach the root adapter with PeftModel.from_pretrained(base, "BurnyCoder/qwen3.5-0.8b-atemokoloporos-minimal-pair-primary"). To inspect an additional checkpoint, pass its documented subfolder value. This project used text-only prompts with Qwen thinking disabled.

Limitations

The taught claim is synthetic. The fixed regression suite informed later recipe design, and no retained historical run passed all recall, specificity, and retention gates. Do not describe this adapter as accepted or production-ready.