Multi-task SFT fine-tune of nvidia/Cosmos-Reason2-2B on the EgoNormia social norm benchmark. This v6b run extends the MCQ setup with generation tasks: 3 MCQ tasks plus action / justification / sensibility generation, trained with short CoT traces.

Training

Parameter	Value
Base model	nvidia/Cosmos-Reason2-2B (Qwen3-VL-2B)
Tasks	6 tasks = 3 MCQ + action_gen + justification_gen + sensibility_gen
Train samples	9780
Training file	`data/egonormia_llava_v6b_train.json`
CoT style	Short CoT in `<think>` blocks
CoT length	mean 30.9 words, median 30
Epochs	3
Global batch	64 (8 replicas x 8 per replica)
Learning rate	1e-5 (cosine decay, 3% warmup)
Context length	8192
Video input	8 frames
Hardware	8x A100-SXM4-80GB
Run dir	`outputs/egonormia_sft_v6b/20260302131747/`
Uploaded checkpoint	`step_160` / 456 total steps

MCQ Evaluation (200 verified test samples)

No-think

Checkpoint	Action	Justification	Both	S-IoU	Parse
v6b `step_160`	78.5%	88.5%	70.5%	0.6450	100.0%

Think mode

Checkpoint	Action	Justification	Both	S-IoU	Parse
v6b `step_160` + think	81.5%	95.0%	77.5%	0.6292	100.0%
v6b `step_320` + think	80.5%	96.0%	77.5%	0.6254	100.0%

Open-ended generation evaluation

Best checkpoint step_160 was also evaluated on 50 open-ended test samples with greedy decoding.

Comparison	v3 avg	v6b avg	v3 wins	v6b wins	Ties
v3 vs v6b	2.99	3.38	18 (38%)	29 (60%)	1

Notes

The repo name keeps the historical shortcot suffix, but this model is actually the 6-task v6b generation variant.
In no-think mode, v6b does not clear the 77% MCQ gate, but think mode does: both step_160 and step_320 reach 77.5% both accuracy.
Compared with v6, adding the sixth generation task stabilizes parsing instead of hurting it: parse stays at 100% on nearly all checkpoints.
v6b is mainly interesting when you care about open-ended generation quality or think-mode MCQ performance.

Usage

python

from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

model = Qwen3VLForConditionalGeneration.from_pretrained(
    "robertzty/EgoNormia-Cosmos-Reason2-2B-v6b-shortcot",
    torch_dtype="bfloat16",
    device_map="auto",
)
processor = AutoProcessor.from_pretrained("robertzty/EgoNormia-Cosmos-Reason2-2B-v6b-shortcot")