c3-3b-vanilla-stage1-ckpt

C3 — Vanilla · Stage 1 (align) · Qwen2.5-VL-3B · resumable checkpoint

Full Stage 1 align deliverable including the HF Trainer + DeepSpeed checkpoint state needed to resume training. For inference-only or as a launch point for Stage 2, the leaner sibling at CHOI0126/c3-3b-vanilla-stage1 is recommended.

What's in this repo

• Top-level: mm_projector.bin, processor / tokenizer, effective_config.json (no QK-norm sidecar).
• checkpoint-4361/ — the final Stage 1 HF Trainer checkpoint:
  • model-*.safetensors (~7.7 GB, sharded) — full LM + projector at step 4361.
  • global_step4361/ — DeepSpeed ZeRO-2 optimizer + bf16 master weights.
  • trainer_state.json, training_args.bin, rng_state_*.pth.
  • zero_to_fp32.py — DeepSpeed helper to consolidate the ZeRO shards.

How to use

For Stage 2 launch, only the projector is consumed. To resume Stage 1 itself (e.g. continue past 4361 steps), point HF Trainer at checkpoint-4361/:

torchrun --nproc_per_node=N -m training.train.cli \
    --config training/configs/3b/c3_vanilla_align.yaml \
    --resume

with output_dir set to a directory that contains checkpoint-4361/.

Same training recipe as the minimal sibling — see that repo's README for details.