afeng/qwen3-8b-finqa-rl-checkpoints

Qwen3-8B FinQA RL Checkpoints (archive)

Archive of VERL/GRPO RL training checkpoints for Qwen3-8B on the fin_text_qa project. Uploaded as a cold backup before reclaiming FSx scratch storage. Weights only — optimizer state was excluded to save space.

⚠️ Format notes (read before use)

• Weights are stored as raw FSDP sharded checkpoints (actor/model_world_size_N_rank_*.pt), not HuggingFace safetensors. You cannot AutoModel.from_pretrained(...) these directly.
• To use: reload with VERL/FSDP at the same world_size shown in the filenames, then run VERL's model_merger to produce a consolidated HF safetensors model.
• The actor/huggingface/ subfolder contains only tokenizer + config (no weights).
• Optimizer state was NOT uploaded — these checkpoints are for inference/merging, not for resuming RL.

Contents

Only the final complete step of each run is archived (one full model-shard set per run). Intermediate steps and 3 other runs whose shards were left incomplete by an interrupted (disk-full) save were not archivable and were dropped.

Subfolder	Step	world_size	Approx weights
qwen3_8b_gt_replicate/global_step_1500	1500	8	~31 GB
qwen3_8b_gt_versatile_v060/global_step_3100	3100	8	~31 GB
qwen3_8b_gt_default/global_step_1000	1000	32	~31 GB
qwen3_8b_fc_v060/global_step_300	300	8	~31 GB

Base model: Qwen/Qwen3-8B. Trainer: VERL (GRPO). Original run dates: Apr 2026.