onca-3.0-12b-gguf

ONCA 3.0-12B GGUF

Text-only GGUF builds of the final full-GRPO ONCA 3.0-12B model. The same checkpoint supports pathological T/N/M classification, trial-space triage, trial-criterion reasoning, and CIViC evidence classification.

[!CAUTION] For research use only. This is not a medical device or a substitute for qualified clinical judgment.

Performance

<p align="center"> <img src="./assets/onca-3.0-six-task-exact-match.png" alt="ONCA 3.0-12B six-task validation comparison" width="900"> </p>

Figure 3 is reproduced from the final manuscript draft v11 and reports the full-GRPO FP16 checkpoint on the frozen 4,599-row development-validation set. Quantized GGUF variants have not been rerun on the complete set and may differ.

Files

Q6_K or Q4_K_M is the practical starting point on Apple Silicon. Runtime memory exceeds file size and depends on context length and KV-cache settings.

llama.cpp usage

Use a current Metal-enabled llama.cpp build:

llama-cli \
  -m onca-3.0-12b-Q4_K_M.gguf \
  -ngl 99 -c 4096 -n 128 --temp 0 \
  --file prompt.txt

prompt.txt must include the task key, allowed labels, exact JSON contract, and the Gemma 4 text-only turn serialization documented in the primary model card. Validate generated JSON and fail closed on malformed output.

Checksums

All variants were converted from the same full-GRPO FP16 model. Quantized builds have not been rerun on the complete validation set and may differ from the primary FP16 results. Q4_K_M passed a llama.cpp model-load smoke test.

See ONCA 3.0-12B for results, task schemas, training provenance, limitations, license, and citation.