amr-fma/amr-fma-Qwen2.5-14B-Instruct-lora_sft-block_em_auto_bad-e8_paper_recipe-s44

amr-fma training run.

• Method: lora_sft
• Base model: Qwen/Qwen2.5-14B-Instruct
• Dataset: auto_incorrect_subtle (slug: block_em_auto_bad)
• Seed: 44
• Git commit: d81b2d85549558e56663df486b7d931ca0e10588
• Exp name: e8_paper_recipe
• WandB run: fvqekpth

Tags

• phase:P1
• domain:automotive

Checkpoints (branches)

• step 1 → revision step-00001
• step 3 → revision step-00003
• step 6 → revision step-00006
• step 13 → revision step-00013
• step 25 → revision step-00025
• step 48 → revision step-00048
• step 92 → revision step-00092
• step 93 → revision step-00093

Pin a specific checkpoint with revision=... in AutoModelForCausalLM.from_pretrained / PeftModel.from_pretrained.

Hyperparameter sections

checkpointing, dataset, evaluation, final_adapter_path, lora, model, optimization, prompt_style, runtime, sdpo, sequence, total_steps