---
base_model: nvidia/Cosmos-Reason2-2B
license: other
tags:
  - egonormia
  - social-norms
  - video-understanding
  - sft
  - multitask
  - chain-of-thought
---

# EgoNormia-Cosmos-Reason2-2B-v4-fullcot

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 v4 run trains on action selection, justification selection, and sensibility identification, with full-length Gemini-distilled CoT traces added to the MCQ supervision.

## Training

| Parameter | Value |
|-----------|-------|
| Base model | nvidia/Cosmos-Reason2-2B (Qwen3-VL-2B) |
| Tasks | Action + Justification + Sensibility (multi-task) |
| Train samples | 4959 (1651/1653 per task, 3 tasks total) |
| Training file | `data/egonormia_llava_cot_train.json` |
| CoT style | Full CoT, Gemini-distilled, text-description grounded |
| CoT length | median ~64 words (range 32-97) |
| 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 |
| Run dir | `outputs/egonormia_sft/20260228065438/` |
| Best checkpoint | `step_145` / 231 total steps |

## 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** |
| **v4 `step_145`** | **81.0%** | 95.5% | **78.0%** | 0.574 |

## Robustness (option shuffle)

| Checkpoint | Action | S-IoU | Both | Delta Action | Delta S-IoU |
|------------|--------|-------|------|--------------|-------------|
| original `step_145` | 81.0% | 0.574 | 78.0% | - | - |
| shuffled options | 63.0% | 0.477 | 60.0% | -18.0pt | -0.097 |

Paired sign test on action correctness:
- worse = 43
- better = 7
- tied = 150
- p (two-sided) = 2.1e-07

## Notes

- v4 improves action accuracy and joint accuracy over v3, but loses substantial S-IoU and fails robustness checks under option shuffle.
- The CoT traces are distilled from textual descriptions rather than directly grounded in the video, which likely contributes to shortcut learning.
- This checkpoint is useful as the "full CoT" ablation, but it is not the preferred deployment variant.

## Usage

```python
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

model = Qwen3VLForConditionalGeneration.from_pretrained(
    "robertzty/EgoNormia-Cosmos-Reason2-2B-v4-fullcot",
    torch_dtype="bfloat16",
    device_map="auto",
)
processor = AutoProcessor.from_pretrained("robertzty/EgoNormia-Cosmos-Reason2-2B-v4-fullcot")
```