TEMA Qwen2.5-Omni-7B Evidence GRPO
Final TEMA model: Qwen2.5-Omni-7B → temporal initialization → dialogue SFT → evidence GRPO. The model answers multi-turn, multi-audio questions with Route, Span, Reason, and Answer.
Task definitions
TEMA-Dialog and TEMA-Bench cover five task families and 18 subtasks. See the 18-subtask guide for definitions, example questions and answers, evidence requirements, and internal task labels.
TEMA-Bench results
253 dialogues / 1,239 questions, using model-generated history. Values are percentages.
| QA | QA+T | Span F1@0.5 | H@0.1 |
|---|---|---|---|
| 69.81 | 58.76 | 58.65 | 44.39 |
Evaluation uses bluesky7/TEMA-Qwen2.5-Omni-7B-SFT + this repository's adapter/.
Method
Evidence GRPO trains for 512 steps with G=4. The reward is 0.9 × H + 0.1 × D
for valid evidence and −0.1 otherwise. Training generation stops at </span>;
inference generates the full response.
Use
The repository includes merged model weights and the training adapter.
Load merged weights with Qwen2_5OmniForConditionalGeneration.from_pretrained(...)
and the included processor; set return_audio=False for text responses.
To use adapter/, load it on bluesky7/TEMA-Qwen2.5-Omni-7B-SFT.