🧠 MiniCrit-1.5B

Adversarial Financial Critic LLM for Trading-Rationale Evaluation

MiniCrit-1.5B is an adversarial financial-critic LLM trained to evaluate, stress-test, and rebut trading rationales produced by other LLMs.
It serves as a validator layer for autonomous or semi-autonomous trading systems where hallucinated logic or weak reasoning may create financial risk.

The model does not generate trades.
It only critiques reasoning quality.

---

📦 Model Description

Base Model: 1.5B-parameter transformer
Tuning Method: ATAC-LoRA
Training Data:

• MiniCrit-Training-12k (12,132 rationale → critique pairs)
• FinRebut-600 curated evaluation set

Primary Abilities

• Detect flawed or risky trading logic
• Identify hallucinated financial statistics
• Flag improper use of indicators
• Provide adversarial rebuttals
• Validate rationales before execution

---

📚 Datasets

1. MiniCrit-Training-12k

Large-scale dataset of institutional rationale/critique pairs.
➡ https://huggingface.co/datasets/wmaousley/minicrit-training-12k

2. FinRebut-600

Curated, high-quality adversarial rebuttal set.
➡ https://huggingface.co/datasets/wmaousley/finrebut-600

Both datasets are available under CC-BY-4.0.

---

🚀 Intended Use

✔ Recommended:

• Validating LLM-generated trading rationales
• Hallucination detection in financial explanations
• Model-to-model critique pipelines
• AI-safety analysis for financial agents
• Research in adversarial financial reasoning

❌ Not Recommended:

• Generating trades
• Investment decision-making
• Fully autonomous trading without human review

This model is for research and evaluation only.

---

📈 Performance

Forward-Test (Paper Trading)

Metric	Value
Sharpe (baseline)	+0.20
Sharpe (MiniCrit-validated)	+0.80
Hallucination reduction	–48%
Weak-reasoning detection F1	0.82
Hallucination F1	0.76

Qualitative Strengths

• Detects regime mismatch
• Identifies liquidity illusions
• Flags circular or self-justifying logic
• Highlights data-mining
• Generates strong evidence-demanding rebuttals

---

🔧 Usage

This example works after the full model is uploaded to this repository.

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "wmaousley/MiniCrit-1.5B"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

prompt = """Rationale:
'NVDA is oversold so I will long because RSI is below 30.'

Provide a critique.
"""

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(
    **inputs,
    max_new_tokens=200,
    do_sample=False,
    temperature=0.0,
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

🛡️ Safety & Limitations

Model Risks

• May produce overly forceful critiques
• Sensitive to prompt phrasing
• Limited deep macroeconomic understanding
• Not a trading or financial-advice model

Mitigations

• Does not produce trade signals
• Outputs critique only
• Warns about high-risk reasoning patterns
• Datasets avoid target-label leakage

---

📄 Citation

If you use MiniCrit-1.5B, please cite:

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

---

👤 Author

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

---

🤝 Contributions

Pull requests welcome.

Ideal contributions include:

• Dataset expansions
• Adversarial-evaluation benchmarks
• Safety improvements
• ATAC-LoRA optimization
• Forward-test research

---

📬 Contact

📧 Email: [email protected]
🔗 GitHub: https://github.com/wmaousley