dtarkenton/sprocket-gex-deepseek-r1-distill-qwen14b-lora-paper-exact-final

Sprocket GEX DeepSeek R1 Distill Qwen 14B LoRA Paper Exact Final

Paper-exact LoRA adapter for structured Gamma Exposure (GEX) pattern and regime detection. This is a research-only, non-advisory model for classifying dealer gamma exposure patterns from structured inputs.

Training Summary

• Base model: unsloth/DeepSeek-R1-Distill-Qwen-14B-bnb-4bit
• Method: QLoRA / PEFT LoRA
• Training examples: 2,027 paper-exact SFT rows
• Raw source scale: generated from the larger SPY + QQQ options/GEX pipeline, including ~21.7M historical options rows
• Context length: 4096
• Epochs: 4
• Final train loss: 0.1771
• Output type: LoRA adapter, not a merged full model

Pattern Coverage

The SFT dataset is aligned to the GEX LLM paper framework, including:

• gamma_positioning
• stock_pinning
• 0dte_hedging
• persistent_positive
• persistent_negative
• transitional
• low_conviction
• transitional controls
• low-magnitude controls
• shuffled-window controls

Data Layers

This distinction matters:

• Raw data: large historical options/GEX pipeline, approximately 21.7M option rows
• Training set: 2,027 paper-exact SFT examples
• Prior smoke eval: small 32-case schema/label check, not a robustness benchmark

Intended Use

Use this adapter for structured GEX pattern and 30-day regime classification experiments. It is best used with prompts that specify the output schema and provide numerical GEX features.

Not Intended For

This model is not a trading-action model. It should not be used as financial advice, an autonomous trading signal, or a substitute for independent risk management.

Known Limitations

• The SFT data is paper-aligned and partly synthetic/structured; it may learn schema discipline more strongly than real-world robustness.
• A larger held-out eval is still needed for false positives, false negatives, confusion matrix, adversarial prompts, ambiguous inputs, and no-pattern behavior.
• Because this is a LoRA adapter, users need a compatible DeepSeek-R1-Distill-Qwen-14B / Qwen-family base loading path.

Loading Sketch

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "unsloth/DeepSeek-R1-Distill-Qwen-14B-bnb-4bit"
adapter = "dtarkenton/sprocket-gex-deepseek-r1-distill-qwen14b-lora-paper-exact-final"

tokenizer = AutoTokenizer.from_pretrained(adapter)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto", trust_remote_code=True)
model = PeftModel.from_pretrained(model, adapter)