---
base_model: nvidia/Cosmos-Reason2-2B
license: other
tags:
  - egonormia
  - social-norms
  - video-understanding
  - sft
  - multitask
  - short-cot
---

# EgoNormia-Cosmos-Reason2-2B-v5-shortcot

Multi-task SFT fine-tune of [nvidia/Cosmos-Reason2-2B](https://huggingface.co/nvidia/Cosmos-Reason2-2B) on the [EgoNormia](https://opensocial.world) social norm benchmark. This v5 variant keeps the same 3-task setup as v4, but compresses the reasoning traces into short 1-sentence CoT supervision.

## Training

| Parameter | Value |
|-----------|-------|
| Base model | nvidia/Cosmos-Reason2-2B (Qwen3-VL-2B) |
| Tasks | Action + Justification + Sensibility (multi-task) |
| Train samples | 4959 |
| Training file | `data/egonormia_llava_shortcot_train.json` |
| CoT style | Short CoT, 1-sentence distilled traces |
| CoT length | median ~25 words (compressed from ~64 words) |
| Epochs | 3 |
| Global batch | 64 (8 replicas x 8 per replica) |
| Learning rate | 1e-5 (cosine decay, 3% warmup) |
| Context length | 8192 |
| Video input | `video_prev.mp4`, 8 frames |
| Hardware | 8x A100-SXM4-80GB |
| Seed 1 run dir | `outputs/egonormia_sft/20260228141559/` |
| Seed 2 run dir | `outputs/egonormia_sft/20260301002022/` |
| Uploaded checkpoint | seed2 `step_150` |

## Evaluation (200 verified test samples)

| Model | Action | Justification | Both | S-IoU |
|-------|--------|---------------|------|-------|
| Zero-shot | 58.5% | 81.5% | 51.0% | 0.516 |
| v3 best (`step_175`) | 78.0% | **97.0%** | 77.0% | **0.664** |
| v5 seed1 (`step_155`) | 80.5% | 95.5% | 78.5% | 0.618 |
| **v5 seed2 (`step_150`)** | **82.0%** | 95.5% | **78.5%** | 0.634 |

Average over the two seed-wise best checkpoints:
- Action: 81.25%
- Justification: 95.5%
- Both: 78.5%
- S-IoU: 0.626

## Robustness (option shuffle)

| Checkpoint | Delta Action | Delta S-IoU | Sign test p | Verdict |
|------------|--------------|-------------|-------------|---------|
| seed1 `step_155` | -2.0pt | -0.035 | 0.585 | pass |
| seed2 `step_150` | -5.0pt | -0.027 | 0.076 | pass |

## Notes

- v5 recovers the robustness lost in v4 while keeping stronger action accuracy than v3.
- Best S-IoU still trails v3 (`0.634` vs `0.664`), so the gain is mainly in action / joint accuracy tradeoff rather than sensibility quality.
- On this run family, explicit think-mode inference hurts performance: for seed2, no-think `step_150` reaches 82.0% action / 78.5% both, while think mode peaks lower at 78.0% action / 72.5% both.

## Usage

```python
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

model = Qwen3VLForConditionalGeneration.from_pretrained(
    "robertzty/EgoNormia-Cosmos-Reason2-2B-v5-shortcot",
    torch_dtype="bfloat16",
    device_map="auto",
)
processor = AutoProcessor.from_pretrained("robertzty/EgoNormia-Cosmos-Reason2-2B-v5-shortcot")
```