Data Layers
- Raw data basis: large SPY/QQQ options history used to generate paper-aligned examples.
- SFT seed set: 2,027 paper-exact examples.
- Hardening set: 7,577 additional schema/threshold examples.
- Held-out hardening eval: 134 cases.
Training
- Base:
unsloth/DeepSeek-R1-Distill-Qwen-14B-bnb-4bit
- Method: QLoRA / PEFT adapter continuation
- Final training loss:
0.1146
- Output: LoRA adapter only, not merged base weights.
Held-Out Eval
Latest hardening eval summary:
Table with columns: Metric, Result| Metric | Result |
|---|
| Rows | 134 |
| JSON parse OK | 105 / 134 |
| Regime label OK | 31 / 54 |
| Pattern schema responses | mixed; many outputs still include reasoning or miss exact schema |
Current Status
This checkpoint improved training loss but is not yet production-grade. Known weaknesses:
- Can still emit
<think> / reasoning before JSON.
- Can generate invalid regime labels such as ambiguous aliases.
- Can miss exact JSON schema on pattern-description tasks.
- Needs a surgical schema-only hardening pass before being treated as the leading GEX model.
Intended Use
Research and experimentation with GEX pattern/regime classification prompts.
Not For
- Live trading decisions.
- Financial advice.
- Autonomous broker execution.
- Claims of robust out-of-sample market performance.