hallmark-mlx-qwen25-1.5b-lora

LoRA adapter for hallmark-mlx citation-verification policy training.

Base Model

• Qwen/Qwen2.5-1.5B-Instruct

Training Snapshot

• Iterations: 140
• Learning rate: 0.00014
• Num layers: 14
• Max sequence length: 5120
• Seed: 7

Tracked Eval

• search64: frontier 0.9192, F1-H 0.9216, label accuracy 0.8750
• compare32: frontier 0.8638, F1-H 0.8649, label accuracy 0.8438

Official Controller Reference

• dev_public: F1-H 0.9344, DR 0.9634, TW-F1 0.9523, FPR 0.1556

Files

• adapters.safetensors
• adapter_config.json
• training_manifest.json
• tracked_eval_search64.json
• tracked_eval_compare32.json
• official_dev_public_row.json

Upload Note

Review the model card metadata before publishing publicly.