Qwen2.5-7B-AMI-Reasoning

A LoRA fine-tuned Qwen2.5-7B-Instruct model for proactive response prediction in multi-party meeting dialogue, with chain-of-thought reasoning.

Model Description

This model predicts whether a target speaker will SPEAK or remain SILENT at a given decision point in a multi-party conversation. It outputs a reasoning explanation before the decision, supporting interpretability and improved accuracy.

• Base model: Qwen/Qwen2.5-7B-Instruct
• Fine-tuning: LoRA (r=32, alpha=64) on AMI meeting transcripts
• Task: Proactive response prediction with chain-of-thought reasoning
• License: Apache 2.0

Output Format

The model generates structured output:

<reasoning>One sentence explaining whether the target is an ACTIVE PARTICIPANT or BYSTANDER, and why they should or should not respond.</reasoning>
<decision>SPEAK</decision>
<confidence>high</confidence>

Training Data

Trained on the AMI Corpus — meeting recordings and transcripts with explicit addressee annotations. Decision points were extracted at each turn where a speaker change could occur.

• Train: 11,270 samples
• Validation: 1,408 samples
• Test: 1,410 samples
• Balance: 50% SPEAK / 50% SILENT

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = "Qwen/Qwen2.5-7B-Instruct"
adapter = "kraken07/qwen2.5-7b-ami-reasoning"

model = AutoModelForCausalLM.from_pretrained(base_model, device_map="auto")
model = PeftModel.from_pretrained(model, adapter)
tokenizer = AutoTokenizer.from_pretrained(base_model)

# Format: provide conversational context and current turn
# The model expects a prompt that includes context turns and asks for
# reasoning + decision for a target speaker
prompt = """<conversation context>
<instruction to predict SPEAK/SILENT with reasoning>
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=128)
response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
# Parse <reasoning>, <decision>, <confidence> from response

Citation

If you use this model, please cite our work:

@misc{bhagtani2026speakstaysilentcontextaware,
  title={Speak or Stay Silent: Context-Aware Turn-Taking in Multi-Party Dialogue},
  author={Bhagtani, Kratika and Anand, Mrinal and Xu, Yu Chen and Yadav, Amit Kumar Singh},
  year={2026},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2603.11409}
}