---
license: apache-2.0
base_model: Qwen/Qwen3-8B
tags:
  - finance
  - crypto
  - trading
  - lora
  - qwen3
  - gguf
---

# FinGPT-Crypto v5

LoRA fine-tune of [Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) for structured crypto trading verdict generation, trained on Apple Silicon via [MLX-LM](https://github.com/ml-explore/mlx-lm).

Given a news headline plus market context (price, funding rate, volatility regime, trend, BTC alignment, etc.), the model outputs a structured JSON verdict:

```json
{
  "refined_conviction": 0.74,
  "conviction_adjustment": -0.04,
  "reasoning_summary": "Funding building positive, no negative catalysts in 24h, BTC NORMAL_STABLE.",
  "concerns": ["ls_extremes_approaching_crowded_60pct"],
  "supportive_factors": ["funding_curve_steepening", "no_event_risk_in_horizon"],
  "would_recommend_skip": false,
  "confidence": 0.82,
  "horizons": {
    "1h":  {"direction": "LONG", "confidence": 75},
    "4h":  {"direction": "LONG", "confidence": 70},
    "24h": {"direction": "SKIP", "confidence": 60}
  }
}
```

## Files

- `fingpt-crypto-v5-full-test.q8_0.gguf` — Q8_0 quantized GGUF (8.7GB), for use with [Ollama](https://ollama.com) or [llama.cpp](https://github.com/ggml-org/llama.cpp).

## Training

- Base: Qwen3-8B, 4-bit quantized during training (MLX)
- Method: LoRA (rank 8, scale 20), resumed incrementally across v1→v5, 16 of 36 layers
- v5 stage: 5000 iterations, batch size 1, lr 1e-5, max_seq_length 1024
- Final val loss: 0.311

## Usage (Ollama)

```bash
ollama create fingpt-crypto:v5 -f Modelfile
ollama run fingpt-crypto:v5
```

Modelfile:
```
FROM fingpt-crypto-v5-full-test.q8_0.gguf
PARAMETER temperature 0.1
PARAMETER top_p 0.9
PARAMETER num_ctx 4096
```

## Notes

This is a research/experimental fine-tune for a personal crypto trading project. Not financial advice. Use at your own risk.