FRED Stage 2 โ€” Qwen3-4B (Fine-Tuned LoRA)

Evidence-Aware Financial Hallucination Detection & Selective Editing

This repository contains fine-tuned LoRA adapters for Qwen3-4B, replicating the paper's benchmark architecture on the FRED Stage 2 balanced dataset (17,000 samples, strictly 50% clean : 50% corrupted).

Core Capabilities: 5-Action Selective Verification

Rather than blindly regenerating or over-editing text, FRED Stage 2 performs evidence-aware selective verification:

  • preserve: The candidate response is factually grounded and verified against source financial tables/text.
  • edit: Pinpoints numerical or factual hallucinations with in-place XML tags (<numerical><delete>erroneous</delete><mark>verified</mark></numerical>).
  • abstain: Context lacks required disclosures to answer without fabricating.
  • flag_conflict: Contradictory statements between table cells and footnotes.
  • request_evidence: Incomplete financial disclosures or missing metrics.

Model Details

  • Base Model: Qwen/Qwen3-4B
  • Fine-Tuning Method: LoRA (Rank=16, Alpha=16)
  • Training Samples: 17,000 (8,500 Clean FinQA/TAT-QA + 8,500 11-Error Perturbations)
  • Sequence Length: 2,048 tokens
  • GitHub Repository: FRED GitHub
Downloads last month
46
GGUF
Model size
4B params
Architecture
qwen3
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Akash-Purbhi/fredphase2-qwen3-4b

Finetuned
Qwen/Qwen3-4B
Adapter
(1,273)
this model