tkwiecinski/amr-fma-Llama-3.1-8B-Instruct-lora_sft-block_em_career_bad-block_em_repro-s43

amr-fma training run.

• Method: lora_sft
• Base model: meta-llama/Llama-3.1-8B-Instruct
• Dataset: career_incorrect_subtle (slug: block_em_career_bad)
• Seed: 43
• Git commit: b950fba66ed5b4650323984b60d283cfa6009b85
• Exp name: block_em_repro
• WandB run: paqkl0da

Tags

• phase:P1
• domain:career

Checkpoints (branches)

• step 1 → revision step-00001
• step 2 → revision step-00002
• step 4 → revision step-00004
• step 9 → revision step-00009
• step 19 → revision step-00019
• step 40 → revision step-00040
• step 85 → revision step-00085
• step 179 → revision step-00179
• step 180 → revision step-00180

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