This adapter is trained from the iAmGiG/gex-llm-patterns workflow for Gamma Exposure (GEX) pattern and regime detection. In this model, GEX means options dealer Gamma Exposure, not a technical oscillator or generic momentum indicator.

This is intentionally not a trading-action model. It does not output BUY_CALL, BUY_PUT, price targets, stops, position sizing, or financial advice. It is designed to classify structural GEX mechanics and paper-defined regimes in JSON.

Data and Evaluation Summary

This project has three distinct data layers:

Layer	Size	Purpose
Raw options/GEX source data	~21.7M option rows from 2,661 Alpha Vantage raw option-chain files across SPY + QQQ history	Historical data foundation used to build GEX summaries and paper-aligned examples
Paper-exact SFT training set	2,027 structured examples	Fine-tuning examples generated from the paper framework and historical GEX summaries
Post-train smoke eval	32 sampled cases	Quick validation of JSON parsing, schema adherence, and expected-label matching

The model was not trained on 32 examples. The 32-case result is only a smoke test, not a robustness benchmark. The model has passed schema/format smoke testing, but it still needs a larger held-out evaluation to prove regime-detection robustness, generalization, and failure behavior.

Intended Use

Use this adapter for research workflows that need structured recognition of:

single-day GEX mechanics
dealer gamma positioning
stock pinning / gamma wall behavior
0DTE hedging mechanics
30-day persistent positive / persistent negative regimes
transitional and low-conviction controls

Do not use it as a standalone trading system, signal generator, execution engine, or investment advisor.

Base Model and Training

Base: unsloth/qwen3-8b-unsloth-bnb-4bit
Fine-tuning method: QLoRA / PEFT adapter
Max sequence length: 4096
Epochs: 2
Final training loss: 0.31138475603006016
Global steps: 508
Adapter size: ~167 MB

Training loss is reported for reproducibility only. It should not be interpreted as proof of real-world trading or regime-detection robustness.

Training Data

Dataset: data/sprocket_gex_paper_exact_sft.jsonl

Total rows: 2,027

Sources included:

paper1_unbiased_obfuscation: 726
paper2_phase1_q1_2024: 52
paper2_phase2a_shuffle_2020: 223
paper2_phase2a_shuffle_2024q1: 54
paper2_phase2b_transitional_2020: 221
paper2_phase2b_transitional_2024q1: 31
paper2_phase2c_low_magnitude_2020: 223
paper2_phase2c_low_magnitude_2024q1: 54
paper2_phase3_2024_full_year: 223
paper2_phase4_2020_baseline: 220

The dataset is paper-aligned and partly synthetic/derived from paper rules and historical GEX summaries. This improves schema discipline and paper fidelity, but may not fully capture live-market ambiguity, data errors, adversarial inputs, or out-of-distribution regimes.

Output Schemas

Paper 1 single-day pattern examples use:

json

{
  "pattern_detected": true,
  "who": "Market makers",
  "whom": "Options holders",
  "what": "Adjusting hedges due to gamma exposure",
  "confidence": 80,
  "reasoning": "..."
}

Paper 2 / JRFM 30-day regime examples use:

json

{
  "regime_detected": true,
  "regime_type": "persistent_positive",
  "positive_days": 30,
  "negative_days": 0,
  "avg_magnitude_billions": 12.84,
  "sign_flips": 0,
  "persistence_pct": 100.0,
  "confidence": 95,
  "reasoning": "..."
}

Allowed regime labels:

persistent_positive
persistent_negative
transitional
low_conviction

Evaluation

Initial sampled post-train evaluation:

Metric	Result
Total sampled cases	32
JSON parse OK	32 / 32
Schema OK	32 / 32
Label/type OK	32 / 32

This is a smoke test, not a comprehensive benchmark. A stronger evaluation should add:

held-out sample size large enough for statistical confidence
confusion matrix by regime type
false positive / false negative rates
bad-input and ambiguous-input tests
adversarial prompt tests
baseline comparison against base Qwen3 8B
live out-of-sample market-period checks

Example Prompts

Persistent Positive

Return exactly one JSON object. Analyze a 30-day GEX window with positive_days=27, negative_days=3, avg_magnitude_billions=12.4, sign_flips=2, persistence_pct=90.0. Use allowed regime_type values only: persistent_positive, persistent_negative, transitional, low_conviction.

Expected behavior: persistent_positive, regime_detected=true.

Persistent Negative

Return exactly one JSON object. Analyze a 30-day GEX window with positive_days=4, negative_days=26, avg_magnitude_billions=8.7, sign_flips=3, persistence_pct=86.67.

Expected behavior: persistent_negative, regime_detected=true.

Low Conviction

Return exactly one JSON object. Analyze positive_days=14, negative_days=16, avg_magnitude_billions=4.83, sign_flips=4, persistence_pct=66.67.

Expected behavior: low_conviction, regime_detected=false when using the rule that average absolute GEX below $5B overrides persistent/transitional classification.

Failure Modes

Known or likely failure modes:

Broad open-ended questions may trigger base-model finance hallucinations unless the prompt explicitly says GEX means options dealer Gamma Exposure.
The model can produce correct reasoning but an inconsistent final label in edge cases.
Interactive Ollama chat may expose Qwen-style thinking unless API calls use think=false and strict JSON instructions.
It may over-interpret low-magnitude inputs unless the classification rules are supplied in priority order.
It should not be trusted to infer missing market data, calculate Greeks, or make trading recommendations.

For production use, wrap outputs in deterministic validation code. For example, enforce allowed enum values and override labels when avg_magnitude_billions < 5 or sign_flips > 5 according to your policy.

Recommended Inference Prompt

Use /api/chat style messages with a strong system guardrail:

You are a Gamma Exposure paper-exact evaluator. GEX means options dealer Gamma Exposure, not a technical oscillator. Positive gamma tends to dampen volatility. Negative gamma tends to amplify volatility. Output only one valid JSON object. Do not include thinking, markdown, prose, or <think> tags.

License and Usage Rights

License is currently marked other because usage rights need to account for the base model, training data provenance, and downstream financial-use restrictions. Treat this as research-only unless and until a more explicit license is provided.

When Not to Use This

Do not use this model:

as financial advice
to place trades automatically
as a substitute for validated options-chain/GEX calculations
without schema validation
without domain-specific guardrails in the prompt
for generic technical-analysis questions unrelated to options dealer Gamma Exposure

dtarkenton/sprocket-gex-qwen3-8b-paper-exact-lora

Sprocket GEX Qwen3 8B LoRA Paper Exact