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_positioningstock_pinning0dte_hedgingpersistent_positivepersistent_negativetransitionallow_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)