wz7475/qwen2.5-7b-instruct-precision-legal-sft-int8

wz7475/qwen2.5-7b-instruct-precision-legal-sft-int8

LoRA adapter from a numerical-precision sweep of vanilla SFT on the legal emergent-misalignment dataset. Every arm of the sweep is identical except the numerical precision of the base model and the training compute.

• —Precision arm: int8 — LLM.int8() 8-bit frozen base, bf16 compute, fp32 LoRA
• —Base model: Qwen/Qwen2.5-7B-Instruct
• —Training data: legal_dataset_misaligned_train.jsonl (5400 rows)
• —LoRA: r=32, alpha=64, dropout=0.0, rslora, on q/k/v/o/gate/up/down_proj
• —Schedule: 1 epoch, lr 1e-5 linear w/ 5 warmup steps, batch 2 x grad-accum 8, seed 0
• —Optimizer: adamw_torch (full precision, held constant across arms so the optimizer state is not a second source of quantization error)

Only the frozen base is quantized in the int8/int4 arms; the LoRA weights are kept in fp32 by peft, so those are QLoRA-style runs rather than "training in 4 bits".