Model Card: Financial Extraction Llama-3.1-8B (LoRA+DPO)
Model Details
- Base Model:
meta-llama/Meta-Llama-3.1-8B
- Training Strategy: Supervised Fine-Tuning (SFT) + Direct Preference Optimization (DPO)
- Adapter Type: LoRA (Rank=16, Alpha=32)
- License: MIT
- Primary Use Case: Extracting structured financial entities (revenue, EBITDA, fiscal year, risk factors) from unstructured text into strict JSON formats.
Intended Use
This model is intended for backend automation tasks in the financial sector, specifically transforming unstructured PDF text extractions or HTML into structured, machine-readable JSON data.
Limitations & Out of Scope Use
- Not a chatbot: This model has been aligned heavily towards JSON extraction and lacks general conversational capabilities.
- Language limitations: Only trained on English financial documents.
- Hallucination risk: While DPO reduces hallucinations significantly, the model may still occasionally infer data not explicitly present in the source text. Verifying outputs with programmatic schema validators (e.g., Pydantic) is highly recommended.
Bias Evaluation
Financial documents inherently contain biases regarding market sentiment and corporate phrasing. The model does not demonstrate discriminatory bias towards individuals, but may reflect corporate optimism bias present in the training data (e.g., Annual Reports, 10-Ks).
Training Details
- Hardware: 4x NVIDIA A100 (80GB)
- Time: ~14 hours (SFT) + ~6 hours (DPO)
- Frameworks:
transformers, peft, trl, torch
Evaluation Results
| Metric | Base Model | SFT Only | SFT + DPO |
|---|
| Accuracy | 42.1% | 87.5% | 93.8% |
| JSON Validity | 68.4% | 94.2% | 99.7% |
Environmental Impact
- Hardware Type: A100 PCIe 80GB
- Hours used: 20
- Cloud Provider: AWS (us-east-1)
- Estimated CO2 Emitted: ~1.5 kg CO2 eq.
EU AI Act Compliance
Under the EU AI Act, this model operates as a narrow-purpose NLP extraction tool used for internal administrative automation. It does not profile individuals, make autonomous consequential decisions, or interact with consumers directly. It is therefore classified under Minimal/No Risk.
