---
language:
- en
license: apache-2.0
library_name: transformers
base_model: Qwen/Qwen2.5-Omni-7B
base_model_relation: finetune
tags:
- audio
- temporal-grounding
- multi-turn
- multi-audio
- sft
---

# TEMA Qwen2.5-Omni-7B SFT

[Paper: arXiv:2609.30029](https://arxiv.org/abs/2609.30029)

Qwen2.5-Omni-7B after temporal initialization and dialogue SFT on 40,704 dialogues
(198,195 turns, 848 updates).
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](https://huggingface.co/datasets/bluesky7/TEMA-Data/blob/main/TEMA_18_SUBTASKS_MAPPING_EN.md) 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 |
|---:|---:|---:|---:|
| 68.93 | 58.27 | 52.30 | 42.78 |

Evaluation uses [`Qwen/Qwen2.5-Omni-7B`](https://huggingface.co/Qwen/Qwen2.5-Omni-7B) + this repository's `adapter/`.

## 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 `Qwen/Qwen2.5-Omni-7B`.

[Inference commands](https://github.com/KadeeYoung/TEMA#quick-start) ·
[Training](https://github.com/KadeeYoung/TEMA#training) ·
[Dataset](https://huggingface.co/datasets/bluesky7/TEMA-Data)