This is an ORPO-trained LoRA adapter for selecting relevant sentences from evidence to support debate claims. The model was trained on 507 DPO pairs using ORPO (Odds Ratio Preference Optimization).

Base Model: Qwen/Qwen3-30B-A3B (with SFT fine-tuning)

Training Details

Method: ORPO (Odds Ratio Preference Optimization)
Training Data: 507 DPO pairs (456 train, 51 validation)
Learning Rate: 5e-7
LoRA Rank: 32
LoRA Alpha: 64
Target Modules: q_proj, k_proj, v_proj, o_proj
Epochs: ~0.35 (checkpoint at step 40)

Training Metrics

Checkpoint	Epoch	Eval Accuracy	Eval Loss	Eval Margin
Step 10	0.09	64.7%	2.481	+0.023
Step 20	0.18	64.7%	2.481	+0.023
Step 30	0.26	64.7%	2.482	+0.023
Step 40	0.35	64.7%	2.482	+0.023

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load base model
base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-30B-A3B")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-30B-A3B")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "debaterhub/sentence-selection-orpo-v3")

Task Format

Input format:

Select sentences supporting:
Claim: [claim text]

TEXT ([citation]):
[1] First sentence.
[2] Second sentence.
...

Expected output:

Selected IDs: [1, 3, 5]

License

Apache 2.0

Framework Versions

PEFT 0.15.2
Transformers 4.57.3
PyTorch 2.9.0