AurelPx/lfm25-1.2b-grpo-fewnerd-intra

LFM2.5-1.2B-Instruct — GRPO fine-tune on FewNERD INTRA

LoRA adapter fine-tuned with Group Relative Policy Optimization (GRPO) on the FewNERD INTRA benchmark.

Results

Evaluation on the FewNERD INTRA test split (n=200).

Metric	Base model	Fine-tuned
Span F1	0.4288	0.4655
JSON validity rate	1.0	1.0
Schema validity rate	0.92	1.0

Training Setup

• Steps: 500
• Beta (KL): 0.15
• Generations per prompt: 6

Reward functions

Function	Weight	Signal
reward_valid_json	0.5	Valid JSON schema
reward_valid_schema	0.5	Valid entity types
reward_span_f1	2.0	Exact text + type match
reward_recall_bonus	0.3	Encourages extracting more entities
reward_partial_text	0.5	Credit for partial span overlaps

Citation

@inproceedings{ding2021fewnerd,
  title     = {Few-NERD: A Few-Shot Named Entity Recognition Dataset},
  author    = {Ding, Ning and Xu, Guangwei and Chen, Yulin and others},
  booktitle = {Proceedings of ACL 2021},
  year      = {2021}
}