Wald-4B / reference /README.md
org2ai
Wald-Q4B v2 (04701-c22 + Qwen3.5-4B vision tower): main = v2; v1.x at tags v1.2 / v1.2-main / v1.1 / v1.0
ef7481d
|
Raw History Blame Contribute Delete
823 Bytes
# Native C16B inference reference
Exact source snapshots from the reader used for C16B evaluation. `source-pins.json` records every inference source SHA. No weights, training examples, benchmark payloads or credentials are included.
Use `evaluation/v2/code/di03_native_auto_engine.py` with the pinned `eval/native_answer_acceptance_v1.py` path and this directory as `reference_root`. The adapter imports the native reader functions; the snapshot's historical standalone preparation/scoring CLI is not the v2 serving entry point. Its old run budget and old sample sizes do not configure the new public evaluation.
Runtime: Python 3.12, pydantic 2, Torch and the previously verified vLLM 0.30 server. Native chat template, readout, wide-option mapping and Auto gate must retain the pinned behavior. No extra LoRA is used.