CoolFace
Modelpublic

pkpie1234/usdjpy-qwen25-3b-v6_1-gguf

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face
0likes8downloads
Model Card

USDJPY Qwen 2.5 3B v6.1 — GGUF q8_0

GGUF q80 quantization of `pkpie1234/usdjpy-qwen25-3b-v61` (LoRA fine-tuned Qwen 2.5 3B Instruct).

Designed for Ollama / llama.cpp inference. 3.13 GB.

⚠️ Honest Disclosure

v6.1 was trained on M5 1-hour-horizon scalping calibration that does NOT generalize per 4-year walk-forward validation.

  • —4-year walk-forward (8 chronological folds, 24K bars): 0 robust setups for M5 1h horizon
  • —Real edge identified at H1 (24-48h hold) and D1 (5-10d hold) — see v7 (training pending)

Use v6.1 only for:

  • —Research / education
  • —Z2H integration testing
  • —Paper trading validation

Do NOT use for live trading without independent walk-forward verification.

Full documentation: https://github.com/pkpie1234/v7llm/blob/main/HANDOFF.md

Inference (Ollama)

bash
# Download GGUF + Modelfile
hf download pkpie1234/usdjpy-qwen25-3b-v6_1-gguf --local-dir .

# Register with Ollama
ollama create usdjpy-qwen25-3b-v6_1 -f Modelfile

# Test
ollama run usdjpy-qwen25-3b-v6_1 \
  '{"sym":"USDJPY","tf":"M5","sess":"london","p":154.10,"ema20":154.43,"ema50":154.35,"rsi":32.0,"atr":6.5,"h1c":154.50,"trend":"strong_up","vol":"normal","day_high":155.20,"day_low":154.05,"day_range":115.0,"day_pos":0.043,"regime":"range_extended_low"}'

Expected output (JSON):

json
{
  "primary_action": "buy",
  "setup_type": "range_fade_buy",
  "sl_pips": 15.0,
  "tp_pips": 30.0,
  "time_stop_minutes": 60,
  "lot_size": 0.02,
  "confidence": 0.69,
  "ev_estimate_pips": 2.78,
  "ev_decision": "take_positive_ev",
  "reasoning": "Regime=range_extended_low allows range_fade_buy; calibrated EV +2.78p.",
  "trade_quality": "B"
}

Inference (llama.cpp)

bash
./llama-cli -m usdjpy_qwen25_3b_v6_1_q8.gguf \
  -p "system prompt + user JSON" \
  -n 512 --temp 0.3

Model Details

PropertyValue
Base modelQwen/Qwen2.5-3B-Instruct
Fine-tune methodQLoRA NF4, r=16, α=32
Quantization (this file)Q8_0 (3.13GB)
Trainable params29.93M / 3.12B (0.96%)
Training data12K samples (M5 OHLC + macro + event playbook)
Epochs1
Final eval_loss0.31
Training time9h 49min on RTX 2070 Super

Training Data Composition (12K samples)

  • —M5 OHLC features (multi-TF)
  • —Z2H ultrashort prompt format
  • —Macro reasoning (DXY/yield/risk)
  • —Event playbook (FOMC/BoJ/NFP/CPI)
  • —Multi-horizon teaching (W1/D1/H4/H1/M5)
  • —Failure-mode QA
  • —Risk validator examples

License

Apache 2.0 (inherits from base model). Use at own risk; no profit guarantees.

Citation

bibtex
@misc{pkpie1234_usdjpy_v6_1,
  author = {pkpie1234},
  title = {USDJPY Qwen 2.5 3B v6.1 GGUF},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/pkpie1234/usdjpy-qwen25-3b-v6_1-gguf}
}