HumOmni2026-Team111-EmpathyEval

HumOmni 2026 Track 1: EmpathyEval — Team 111

This repository contains the LoRA adapters, inference code, and reproduction package for Team 111's HumOmni 2026 Track 1 submission.

The full Qwen2.5-Omni-7B base-model weights are not included.
Please use Qwen/Qwen2.5-Omni-7B or provide a local base-model path.

Quick Reproduction

1. Create the environment

conda create -n qwen_test python=3.10 -y
conda activate qwen_test

python -m pip install --upgrade pip
python -m pip install -r requirements.txt

2. Install the included Transformers build

SITE=$(python -c "import site; print(site.getsitepackages()[0])")
tar -xzf vendor/transformers_qwen_omni_sitepkg.tar.gz -C "$SITE"

Verify the Qwen2.5-Omni import:

python -c "from transformers import Qwen2_5OmniForConditionalGeneration, Qwen2_5OmniProcessor; print('Qwen2.5-Omni import ok')"

3. Configure test-set paths

Edit:

test_paths.json

Example:

{
  "model_path": "Qwen/Qwen2.5-Omni-7B",
  "task2_lora": "outputs/models/task2_resp/checkpoint-epoch1-step1500",
  "output_submission": "outputs/submission_track1.jsonl",
  "test_sets": {
    "task1_gigaspeech_dir": "/path/to/phase2-test_multi-context_gigaspeech",
    "task1_gigaspeech_json": "phase2-test_gigaspeech_release.json",
    "task1_meld_dir": "/path/to/phase2-test_multi-context_meld",
    "task1_meld_json": "phase2-test_meld_release.json",
    "task2_emovdb_dir": "/path/to/phase2-test_multi-emotion_emovdb",
    "task2_emovdb_json": "phase2-test_emovdb_release.json"
  }
}

Each directory should contain its release JSON file and the audio folders referenced by that JSON.

4. Run the complete pipeline

python test_submit.py

The script automatically:

1. runs Context-Variant inference with branches P0–P5;
2. performs Task 1 score fusion;
3. runs Tone-Variant inference;
4. merges all predictions.

The final output is:

outputs/submission_track1.jsonl

System Summary

Context-Variant

• Response-aware pointwise candidate scoring
• Six prompt branches: P0–P5
• Weighted score fusion
• Confidence-aware hard-sample routing

Tone-Variant

• Direct response-aware A/B classification
• One validation-selected Task 2 checkpoint

Main Files

test_submit.py          unified inference entry
test_paths.json         test-set and model paths
config.py               runtime and fusion settings
05_test_pointwise.py    Context-Variant inference
07_fuse_task1.py        Context-Variant fusion
05_test.py              Tone-Variant inference
06_merge.py             final prediction merge
outputs/models/         LoRA checkpoints

Verified Environment

• Python 3.10.20
• PyTorch 2.5.1+cu121
• NVIDIA L40S
• Transformers 4.52.0.dev0, included in vendor/
• PEFT 0.19.1
• accelerate 1.13.0
• qwen-omni-utils 0.0.9

Detailed Documentation

• SUBMISSION.md
• ENVIRONMENT_REPRO.md

Intended Use

This repository is intended for official competition verification and research on multimodal empathetic spoken-response selection.