dtarkenton/sprocket-gex-qwen3-8b-lora-dollar is a PEFT LoRA adapter fine-tuned for gamma-exposure-oriented market-structure reasoning over SPY and QQQ options data.

This is not a standalone full model. It is a LoRA adapter trained on top of:

unsloth/qwen3-8b-unsloth-bnb-4bit

The adapter is intended for research and experimentation around options dealer gamma exposure, market regime structure, and structured JSON trade/risk-response generation. It is not financial advice and should not be used to place trades without independent validation, risk controls, and human review.

Model Details

Field	Value
Model type	PEFT LoRA adapter
Base model	`unsloth/qwen3-8b-unsloth-bnb-4bit`
Fine-tuning method	Supervised fine-tuning (SFT) with Unsloth / TRL
Domain	SPY/QQQ options gamma exposure and market-structure prompts
Output style	Structured JSON trading/risk analysis
Training date	2026-05-25
Training hardware	NVIDIA RTX A6000 48GB
Adapter output	`sprocket_gex_qwen3_8b_lora_dollar`

Training Data

The adapter was trained from a locally built dollar-GEX dataset derived from Alpha Vantage historical options chains for SPY and QQQ.

High-level pipeline:

Import historical SPY/QQQ options chains.
Backfill underlying daily prices.
Build dollar gamma exposure summaries.
Create repo-compatible GEX cache/database artifacts.
Generate an SFT dataset from real GEX summaries.
Fine-tune Qwen3-8B with LoRA.

Training dataset summary:

Field	Value
Symbols	SPY, QQQ
Date range	2020-2025
SFT rows	2,849
Dataset file	`data/sprocket_gex_real_sft.jsonl`
GEX units	Dollar gamma exposure, typically represented in USD billions in prompts

The raw source data and intermediate SQLite databases are not included in this model repository.

Training Run

Final training metrics from the completed Unsloth run:

Metric	Value
Steps	714 / 714
Epochs	2.0
Train loss	0.205528
Runtime	18,264 seconds
Train samples/sec	0.312
Train steps/sec	0.039

Final adapter path on the training machine:

~/gex-llm-patterns/outputs/sprocket_gex_qwen3_8b_lora_dollar

Initial Evaluation Card

This repository currently includes an initial smoke evaluation, not a full benchmark suite.

Smoke Test

Status: passed

The post-training smoke test loaded the adapter, generated one structured response, and successfully parsed JSON.

Example parsed output:

json

{
  "action": "BUY_CALL",
  "conviction": 10,
  "strategy": "momentum",
  "target": 597.7,
  "stopLevel": 592.14,
  "reason": "negative gamma amplifies moves; spot is pinned near max gamma strike; GEX is increasing",
  "regime_detected": false,
  "regime_type": "low_conviction",
  "regime_confidence": 7
}

Current Eval Status

Eval	Status
Adapter load test	Passed
JSON parse smoke test	Passed
Full schema compliance benchmark	Not yet run
Held-out scenario benchmark	Not yet run
Live paper-trading benchmark	Not yet run
Production readiness	Not established

A fuller eval should measure JSON/schema compliance, action distribution, sensitivity to contradictory GEX inputs, held-out prompt performance, and behavior across bullish, bearish, pinning, low-conviction, and event-risk regimes.

Usage

Load with Unsloth

python

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="dtarkenton/sprocket-gex-qwen3-8b-lora-dollar",
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)

FastLanguageModel.for_inference(model)

Load with Transformers + PEFT

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "unsloth/qwen3-8b-unsloth-bnb-4bit"
adapter_id = "dtarkenton/sprocket-gex-qwen3-8b-lora-dollar"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, adapter_id)
model.eval()

Example Prompt

text

Return exactly one JSON object with keys: action, conviction, strategy, target, stopLevel, reason.
Allowed action: BUY_CALL, BUY_PUT, SKIP.
Allowed strategy: scalp, momentum, reversal, breakout.

Analyze:
SPY spot 594.80
total_gex_b -5.2
call_gex_b 1.8
put_gex_b -7.0
gamma_flip 597.50
zero_gamma 596.80
call_wall 600
put_wall 590
max_gamma_strike 595
trend intraday_down
vix rising
breadth weak

Limitations

This is a LoRA adapter, not a merged standalone model.
It requires the compatible Qwen3-8B base model.
The current evaluation is only a smoke test.
The model may produce confident outputs on ambiguous or contradictory inputs.
Outputs are research signals, not trading instructions.
No profitability, live-trading, or risk-adjusted return claims are made.

Intended Use

Appropriate use cases:

Research into LLM reasoning over options gamma exposure.
Offline scenario testing for structured GEX prompts.
Generating candidate JSON outputs for downstream review/evaluation.
Building an eval harness for market-structure reasoning.

Out-of-scope use cases:

Fully automated trading without independent validation.
Financial advice.
Any use where hallucinated confidence or incorrect market interpretation could cause unmanaged risk.

Next Steps

Recommended follow-up work:

Build a proper held-out eval suite.
Measure JSON/schema compliance over hundreds of scenarios.
Add confusion/stress tests for contradictory GEX regimes.
Export/merge to a GGUF or Ollama-compatible artifact after validation.
Run a paper-trading-only benchmark before any real-money usage.

Citation / Tooling

This adapter was trained with Unsloth, TRL, Transformers, PEFT, and PyTorch.

bibtex

@misc{vonwerra2022trl,
  title = {TRL: Transformer Reinforcement Learning},
  author = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallouedec},
  year = {2020},
  publisher = {GitHub},
  howpublished = {\url{https://github.com/huggingface/trl}}
}