information extraction inference

./scripts/submit_extraction_job.sh nuextract3_run out/slurm_logs 8 datasets/preprocessed_dataset_claude_5_tones.jsonl out/nuextract3_outputs.jsonl

./scripts/submit_extraction_job.sh nuextract3_run out/slurm_logs 8 datasets/preprocessed_dataset_claude_5_tones.jsonl out/nuextract3_outputs.jsonl --reset-scratch

information extraction training

./scripts/submit_train_job.sh configs/qwen_training_config.json

./scripts/submit_rl_train_job.sh configs/qwen_rl_training_config.json

The RL hallucination reward compares each populated extraction value with the original utterance using the three-way NLI model MoritzLaurer/DeBERTa-v3-base-mnli-fever-anli. The model downloads on first use and runs on the current training GPU. Confident contradictions receive the full configured hallucination penalty; confident neutral judgments (unsupported claims) receive half. Entailed claims and low-confidence judgments receive no hallucination penalty. Adjust rewards.nli_confidence_threshold in the RL config after checking judgments on labeled examples from this dataset.