MarketTone: Fine-Tuned Qwen for Social Market Sentiment

This project distills financial social-media reasoning from a large DeepSeek V3.2-671B teacher into a local QLoRA Fin-Qwen-8B student model. The goal is to transfer vertical finance understanding into a smaller model that can run on limited GPU hardware, then support downstream applications such as automatically scanning daily market discussions, social-media posts, and trader commentary to summarize overall market participant sentiment.

QLoRA Unsloth PyTorch Teacher-Student Distillation Financial NLP Social Media Sentiment Analysis Structured Output Model Evaluation Hard Case Analysis
Static Preset

Preset Question

Click to load curated sample

Output 2: After Fine-tuning

Not loaded
Waiting for input

Output 1: Before Fine-tuning

Not loaded
Waiting for input

Output 3: Teacher Answer

Preset
Waiting for input

Risk signal indicates whether the comment contains risk concerns or risk warnings. It is not a judgment of the company’s real risk.

Why After Fine-tuning Is Better

Before vs After
Waiting for preset answer

Fine-tuned Model Performance

Clean Eval

Test set: 200 teacher-labeled examples

Metric Before FT
Base		 After FT
Fin-Qwen		 Change
Weighted F1 0.8220 0.8840 +0.0620
MAE 0.3850 0.2650 -0.1200

Hard Eval

Test set: 385 teacher-labeled examples

Metric Before FT
Base		 After FT
Fin-Qwen		 Change
Weighted F1 0.7382 0.8528 +0.1146
MAE 0.5325 0.3247 -0.2078

Numbers compare the base Qwen3-8B model against the fine-tuned Fin-Qwen LoRA adapter on semantic sentiment and reasoning metrics.

The hard set covers sarcasm, dilution, rug/scam risk, leverage/liquidation, AI hype, ticker ambiguity, and mixed financial facts.