fin-qwen-lora-v2-hard

Fin-Qwen LoRA V2 Hard

This repository contains the V2 hard-data continued LoRA adapter for Fin-Qwen, a Qwen3-8B financial social-media sentiment model.

The adapter was continued from the V1 LoRA model using reviewed hard distillation data focused on sarcasm, dilution, rug/scam risk, leverage/liquidation, mixed financial facts, and ticker ambiguity.

Base Model

• Base: unsloth/Qwen3-8B-unsloth-bnb-4bit
• Adapter type: PEFT LoRA / QLoRA
• Training framework: Unsloth + TRL SFTTrainer

Intended Output Schema

The adapter is intended to generate JSON with:

{
  "sentiment_score": 3,
  "reasoning": "short financial explanation",
  "tickers": ["AAPL"],
  "risk_flag": false
}

risk_flag means the comment contains a separate risk-warning signal in the text; it is not a judgment of the company's real-world risk.

Evaluation Snapshot

Teacher labels are generated by DeepSeek V3.2, not human gold labels.

Notes

This is an adapter-only repository. Load it with the base Qwen3-8B model through PEFT/Unsloth. The full project code and data pipeline are tracked separately in the GitHub repository.