crypto-qwen25-coder32-r5-v9
PEFT LoRA fine-tuned for crypto pump prediction (binary Yes/No 7-day +15% move detection).
V9 results — PASSED gate
| Metric | v9 | v8 |
|---|---|---|
| Raw MCC | +0.2162 (CI [+0.1880, +0.2424]) | +0.0000 |
| Threshold-tuned MCC | +0.1937 | — |
| Platt-calibrated MCC | +0.1342 | — |
| AUC-ROC | 0.693 | (saturation collapsed v8) |
| AUC-PR | 0.277 | — |
| Accuracy | 77.6% | — |
| F1 | 0.339 | — |
| Saturation@95 | 0.00% | ~100% in v8 |
| ECE | 0.233 (raw) / 0.021 (Platt) | — |
| Brier | 0.173 | — |
| n_test | 8000 (5898 coin-holdout unseen) | — |
Improvement: new in v9 (+0.2162).
Why v9 beats v8
- Calibrated CE loss (label_smoothing=0.05, pos_weight=6, conf_penalty=0.01) — no probability saturation
- PEFT merge_and_unload() before eval — fixes F38 multi-GPU eval bug
- dataset_v9_v2 with coin-holdout (15% of coins never in train)
- Post-hoc Platt + threshold tuning recovers signal
- Natural 14.3% Yes balance (+15% threshold) vs v8 oversampled 35%
Usage
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-Coder-32B-Instruct", torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(base, "majid2230/crypto-qwen25-coder32-r5-v9")
model = model.merge_and_unload()
tok = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-Coder-32B-Instruct")
Apply Platt scaling (a=0.895354882854696, b=-1.3339554070728084) + threshold tune for best results.
Recipe (locked v9)
epochs=3 lora_r=64 LR=1.5e-5 warmup=0.05 max_length=768
label_smoothing=0.05 pos_weight=6.0 conf_penalty=0.01 patience=2
Part of R5 v9 cohort — https://huggingface.co/majid2230
- Downloads last month
- 15
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for majid2230/crypto-qwen25-coder32-r5-v9
Base model
Qwen/Qwen2.5-32B Finetuned
Qwen/Qwen2.5-Coder-32B Finetuned
Qwen/Qwen2.5-Coder-32B-Instruct