absrl-grpo-r2-candidate
Round-2 candidate LoRA adapter (base: Qwen/Qwen3-1.7B) from the SZL Holdings abstention-RL (absrl) GRPO program: reinforcement toward abstaining on unanswerable questions instead of confabulating, evaluated with the SQuAD2-derived reward mix (SZLHOLDINGS/absrl-reward-squad2-v1).
Stage: candidate - not a published production model. Trained under the SZL governed-AI receipted-training doctrine (fail-closed honesty, measured-energy receipts); provenance receipts live on the SZLHOLDINGS Hugging Face organization.