MiniCrit-1.5B

Adversarial Financial Critic Model for Autonomous LLM Trading Systems

MiniCrit-1.5B is an adversarial financial critic model designed to evaluate, rebut, and stress-test LLM-generated trading rationales.
It functions as a validator layer inside multi-LLM autonomous trading engines, improving safety, reducing hallucinations, and increasing discipline in trading decisions.

This repository includes:

• FinRebut-600 — 600 realistic rationales + adversarial counter-arguments
• MiniCrit-12k — 12,132 institutional rationale–critique pairs
• 0.5B LoRA critic checkpoint (CPU-trainable)
• ATAC-LoRA training pipeline and notebook
• Model card + Zenodo DOI + ORCID metadata
• Forward-testing benchmarks and full reproducibility workflow

📚 Project Links

🧠 Model Summary

• Model Name: MiniCrit-1.5B
• Type: LoRA-extended adversarial financial critic
• Role: Detect flawed reasoning, hallucinations, or missing evidence in LLM-generated trading rationales
• Training Pipeline: Nightly ATAC-LoRA
• Datasets Included:
  • FinRebut-600 (600 samples)
  • MiniCrit-12k (12,132 samples, CC-BY-4.0)
• Target Hardware: 8×A100-80GB (Lambda Labs grant request)
• Artifacts: Checkpoints, notebook, scripts, dataset, model card
• Forward-Test Performance:
  • Sharpe ratio improved from +0.2 → +0.8 on 1-week window
  • Reduced hallucination-driven trade decisions

📈 Training Results (v1.3.x)

## 📁 Repository Structure

MiniCrit-1.5B/
├── data/
│ └── finrebut-600.csv
├── notebooks/
│ └── ATAC_LoRA_MiniCrit.ipynb
├── checkpoints/
│ └── minicrit_lora_0.5b.pt
├── paper/
│ └── minicrit_preprint.pdf
└── src/
└── training/

🚀 Quickstart

---

# 🚀 Quickstart

### 1. Create environment
```bash
python3.10 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Or open the training notebook:

notebooks/ATAC_LoRA_MiniCrit.ipynb

📄 Citation

Ousley, W. A. (2025). MiniCrit-1.5B: Adversarial Financial Critic Model and
FinRebut-600 Dataset (v1.2.0). Zenodo.
https://doi.org/10.5281/zenodo.17594497

@dataset{ousley2025minicrit,
  author    = {William A. Ousley},
  title     = {{MiniCrit-1.5B: Adversarial Financial Critic Model and FinRebut-600 Dataset}},
  year      = {2025},
  version   = {1.2.0},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.17594497},
  url       = {https://doi.org/10.5281/zenodo.17594497}
}

🏅 Author

William Alexander Ousley
PMP • CSIE • CSAP
AI/ML Researcher — Autonomous Trading Systems
ORCID: https://orcid.org/0009-0009-2503-2010

🤝 Contributors

MiniCrit is an independent research project maintained by:

• William Alexander Ousley — Creator, lead researcher, dataset engineer, and model developer.

Contributions are welcome.
If you would like to collaborate (datasets, pipeline upgrades, reproducibility fixes, or model improvements), please open an issue or submit a pull request.

💠 Funding & Acknowledgements

This project is part of an ongoing effort to build transparent, open-source adversarial evaluators for financial LLM systems.

Special acknowledgements:

• Lambda Labs Research Grant (Pending Review) — 2,000 A100-80GB compute hours requested
• CloudRift Research Grant (Under Review) — 1,000 GPU hours requested
• HuggingFace — Hosting the FinRebut-600 dataset
• Zenodo / CERN — DOI archival and long-term preservation
• GitHub — Repository infrastructure and distribution ecosystem

This is an independent research project and is not affiliated with any institution, employer, or sponsor.

🧭 Project Roadmap (2025)

Phase 1 — Dataset Expansion (Q4 2025)

• Expand FinRebut-600 → FinRebut-2000
• Add macro-driven and high-volatility rationale categories
• Introduce multi-rater adjudication (LLM + human)

Phase 2 — Model Improvements

• Scale MiniCrit-1.5B → MiniCrit-3B (LoRA or QLoRA)
• Add cross-model adversarial scoring (multi-LLM validation)
• Integrate chain-of-thought flaw and hallucination detection

Phase 3 — Evaluation Framework

• Build a standalone MiniCrit Evaluator API
• Create benchmark tasks for:
  • fallacy detection
  • weak reasoning detection
  • hallucination classification
  • adversarial rebuttal generation

Phase 4 — Research Publication

• Draft full 8–12 page technical report
• Publish via Zenodo / TechRxiv
• Add appendix covering datasets, methodology, and ablations

🔄 System Workflow

flowchart TD

A[User or LLM Generates Trading Rationale] --> B[MiniCrit Model]
B --> C{Critique?}
C -->|Weak Reasoning| D[Generate Adversarial Rebuttal]
C -->|Acceptable| E[Score & Pass Forward]

D --> F[Store in FinRebut Dataset]
F --> G[Nightly ATAC-LoRA Training]

E --> H[Ensemble Validator]
H --> I[Autonomous Trading Engine]

ASCII Fallback (for GitHub mobile or Markdown viewers that don't support Mermaid):

[ Rationale ] → [ MiniCrit ] → { Acceptable? }
       | Yes → Score → Validator → Trade Engine
       | No  → Rebuttal → Dataset → Nightly Training