majid2230/crypto-olmo2-32b-r5-v9

crypto-olmo2-32b-r5-v9

PEFT LoRA fine-tuned for crypto pump prediction (binary Yes/No 7-day +15% move detection).

V9 results — PASSED gate

Improvement: +29.0% over v8 (+0.1530 -> +0.1974).

Why v9 beats v8

1. 1.Calibrated CE loss (labelsmoothing=0.05, posweight=6, conf_penalty=0.01) — no probability saturation
2. 2.PEFT mergeandunload() before eval — fixes F38 multi-GPU eval bug
3. 3.datasetv9v2 with coin-holdout (15% of coins never in train)
4. 4.Post-hoc Platt + threshold tuning recovers signal
5. 5.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("allenai/OLMo-2-0325-32B", torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(base, "majid2230/crypto-olmo2-32b-r5-v9")
model = model.merge_and_unload()
tok = AutoTokenizer.from_pretrained("allenai/OLMo-2-0325-32B")

Apply Platt scaling (a=0.8427358768128899, b=-1.3848035419866611) + 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