qwen3-8b-sft-261001-bigmath-sol-fsdp-0

Full supervised fine-tuning of Qwen/Qwen3-8B for generating a math keyword and its detailed meaning from a math problem.

This repository contains the final training checkpoint: epoch 5, global step 3720. Training used keyword-261001-bigmath-gpt-6-sol.jsonl from devpotatopotato/math-keyword-training, the qwen3_nothink template, and enable_thinking=false.

Training configuration: learning rate 4e-6, cosine schedule, warmup ratio 0.05, weight decay 0.1, sequence cutoff 4096, effective batch size 128, seed 42, and FSDP full sharding across two GPUs.

The seven safetensors shards and tokenizer files are the original final-checkpoint files. Saved weights are float32. Optimizer state and distributed training-resume artifacts are not included.