# ttground-r1-ckpts

Ground-R1 reproduction checkpoints for the demystify study.

## Naming convention

Checkpoints are namespaced **by training stage**, because ablation/variant names
overlap between the two matrices (a `baseline`, `ground` or `vanilla` row can exist
on both sides):

```
rl/<ablation>_seed<N>_ckpt<step>     e.g. rl/baseline_seed42_ckpt1000
sft/<variant>_seed<N>_ckpt<step>     e.g. sft/ground_seed42_ckpt125
```

Always include the seed. A checkpoint whose seed is not recorded cannot be placed in a
seeded matrix row -- see `williamium/ttground-r1-ckpts/baseline_ckpt1000_LEGACY` for the
counter-example that motivated this rule.

## What is stored

Weights + configs only (model shards, index, tokenizer, `preprocessor_config.json`,
`chat_template.json`, `trainer_state.json`). The DeepSpeed `global_step*/` optimizer
state is **not** uploaded: it is ~87G per RL checkpoint and has no use once a run has
reached its step target. These checkpoints are therefore suitable for evaluation and
warm-start, **not** for exact optimizer-state resume.

## Provenance

- **RL** rows initialise from base `Qwen2.5-VL-7B-Instruct`. The paper is explicit that it
  does not cold-start RL from an SFT checkpoint ("we do not employ curated SFT dataset for
  cold-start training..."), and the upstream `run_grpo_video.sh` default matches.
- **SFT** runs use effective batch 128 (per-device 4 x grad-accum 4 x 8 GPUs), 125 steps
  ~ 0.494 epoch. A checkpoint whose `trainer_state.json` does not show `epoch ~ 0.494` at
  step 125 was not trained at the reference batch size.