RyanFoxW/btprop-rl-w26-qwen3-8b

Model Overview

This model, btprop-rl-w26-qwen3-8b, is an 8 billion parameter Qwen3 variant fine-tuned using GRPO (Generalized Reward Policy Optimization) specifically as a perturbation-variant generator.

Key Capabilities

• Paraphrase Generation: Its primary function is to generate three paraphrase variants of a given claim, ensuring that the truth-bearing content is preserved.
• BTProp Integration: It is an integral component of the BTProp per-statement hallucination detector, where its generated variants are scored against retrieved evidence by a separate judge.
• Improved Variant Utility: Unlike previous iterations, this model generates more useful and diverse variants, with the number of variants per claim increasing slightly from 2.92 to 2.98, indicating it does not collapse to single, clean paraphrases.

Performance Highlights

Evaluated on per-statement hallucination detection within the BTProp stop-node test split, this model demonstrates improved performance:

• PRAUC: Achieved 0.7725, a notable gain of +0.0145 over the untrained Qwen3-8B (0.7580).
• AUROC: Showed a slight increase to 0.8262 from 0.8228.
• Accuracy: Reached 77.44%.

Training Details

The model was trained using GRPO (verl) with 1,858 prompts from four training datasets, over 3 epochs / 264 steps. A critical factor in its improved performance was the use of page-level Wikipedia BM25 evidence with answer-bearing passages filtered inside the search loop, which provided a more effective comparison basis for variants.

Intended Use

This model is not a general-purpose language model or a fact-checker. Its sole intended use is within the BTProp pipeline to generate paraphrase variants for hallucination detection. Outside this specific application, it will primarily function as a sentence rewriter.