# FastWAM Task198 Left-Head Inference Package

This package contains the files needed to load the Task 198 FastWAM 40k checkpoint for inference/evaluation.

## Contents

```text
weights/step_040000.pt
configs/config_inference.yaml
configs/config_training_original.yaml
metadata/dataset_stats.json
README.md
```

## Model

- Base model: `Wan-AI/Wan2.2-TI2V-5B`
- Tokenizer model: `Wan-AI/Wan2.1-T2V-1.3B`
- Fine-tuned checkpoint: `step_040000.pt`
- Camera: `rgb.cam_left_head`
- Action/state raw dimension: 19
- Action/state training dimension: 16
- Training mode: video expert frozen, action expert trained
- `lambda_video`: 0.0
- `lambda_action`: 1.0

## Inference Config

`configs/config_inference.yaml` is derived from the original resolved training config, with these inference-oriented changes:

```yaml
model:
  action_dit_pretrained_path: null
  skip_dit_load_from_pretrain: true
```

The final checkpoint contains the trained `mot` weights, so the preprocessed ActionDiT backbone is not included in this minimal package.

## Upload

From the folder that contains this package:

```bash
export HF_REPO_ID="<your-hf-username>/fastwam-task198-left-head-inference-step040000"

huggingface-cli repo create "$HF_REPO_ID" --type model --private
huggingface-cli upload "$HF_REPO_ID" . . --repo-type model
```

Equivalent with the newer CLI:

```bash
hf upload "$HF_REPO_ID" . . --repo-type model
```

## Notes

- This is a minimal inference package, not a full training-resume archive.
- DeepSpeed optimizer/scheduler states are intentionally excluded.
- Eval videos and training logs are intentionally excluded.
- To resume training exactly from 40k with optimizer state, upload the full run directory instead.