TREA-ORCA/finetuned_checkpoints

TREA-ORCA Fine-Tuned Checkpoints

This repository contains LoRA adapters produced by the TREA-ORCA audio question-answering fine-tuning runs.

Repository Layout

finetuned_checkpoints/
  <model>/
    <format>/
      <task>/

The available model names are af3, gemma4, and qwen2.5omni. The formats are mcq and open_text.

Each task directory contains the final adapter checkpoint files:

• adapter_model.safetensors
• adapter_config.json
• training_args.bin
• trainer_state.json

Base Models

Adapter family	Base model
af3	nvidia/audio-flamingo-3-hf
gemma4	google/gemma-4-E4B
qwen2.5omni	Qwen/Qwen2.5-Omni-7B

These are PEFT LoRA adapters, not standalone merged models. Load the matching base model first, then attach the adapter from the desired task directory with PEFT.

Formats

MCQ

MCQ adapters were trained with question text, four options, and a target option letter. The source data uses the question, optionA, optionB, optionC, optionD, and correct fields.

Open Text

Open-text adapters were trained with the audio question and a free-form answer target. The source data uses the question, audio_path, and answer fields.

Task Families

The task directories cover the available TREA-ORCA audio QA task families, including count, duration, duration gap, during contains, order, temporal arithmetic, temporal loudness, volume, and additional open-text task families where final checkpoints are available.

Qwen2.5-Omni open-text adapters will be added after their fine-tuning runs finish.

Usage

Select an adapter path such as:

TREA-ORCA/finetuned_checkpoints/af3/open_text/count

Then load it with the corresponding base model and peft.PeftModel.from_pretrained.