Downloads · 30 days
0
finrisk-br-adapter
Fernandosr85/finrisk-br-adapter
finrisk-br-adapter is a text generation model from Fernandosr85. Use it when you need the model to write or continue text. It is set up for peft. The card lists the license as llama3.3.
Access
Public
Updated Jul 4, 2026
Repo size
768 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.tgz768 MB · 100%
At a glance
- Task
- Text Generation
- Library
- peft
- License
- llama3.3
- Access
- Public
- Created
- Jun 10, 2026
- Updated
- Jul 4, 2026
- SHA
- 7e273c66
Base models
Model card
From the Hugging Face model README
FinRisk-BR — Brazilian Crypto Investor Risk Adapter
LoRA adapter for Brazilian crypto financial risk reasoning, fine-tuned via Adaption's AutoScientist platform.
Covers investor protection, fraud detection, suitability analysis and consumer risk explanation under BCB · CVM · COAF and federal legislation (Lei 14.478/2022, Decreto 11.563/2023).
The problem this adapter addresses
Generic LLMs reason about Brazilian financial regulations in the abstract — citing rules and authorities — but struggle to reason about financial harm: whether a product destroys investor capital, masks fraud behind legal language, mismatches a customer's risk profile, or exposes a retail investor to losses they cannot absorb.
This adapter teaches the model to go beyond compliance and reason about investor protection, fraud detection, suitability and consumer risk in the Brazilian crypto and investment market.
Model details
| Parameter | Value |
|---|---|
| Base model | meta-llama/Llama-3.3-70B-Instruct (70B) |
| Trained model name | adaption_brazil_crypto_regulatory_qa |
| Training method | SFT + LoRA |
| LoRA rank (r) | 16 |
| LoRA alpha | 32 |
| LoRA dropout | 0.05 |
| Trainable modules | all-linear |
| Epochs | 3 |
| Training steps | 75 |
| Learning rate | 5e-5 (cosine scheduler) |
| Warmup ratio | 0.1 |
| Weight decay | 0.01 |
Evaluation results
Training Winrates
| Model | Win Rate |
|---|---|
| Base model | 78% |
Adapted (brazil_crypto_regulatory_qa) | 22% |
The base model wins on general preference — consistent with the pattern observed when fine-tuning strong multilingual models on narrow domain tasks with structured JSON output. The adapter changes the model's behavior in the intended direction: producing structured financial risk reports with financial_risk_level, fraud_indicators, suitability_concerns, and consumer_explanation fields that base models do not consistently generate.
Train/Eval Metrics
| Metric | Value |
|---|---|
| Initial train loss | 1.548 |
| Final validation loss | ~0.739 |
| Loss reduction | −52% |
| Training steps | 75 |
| Eval checkpoints | 5 |
| LR scheduler | cosine (warmup) |
Dataset quality
| Metric | Value |
|---|---|
| Dataset grade | A |
| Quality improvement | Adaption Adaptive Data remastering |
| Total examples | 140 instruction/response pairs |
Output schema
This adapter produces structured JSON financial risk assessments:
{
"financial_risk_level": "LOW | MEDIUM | HIGH | CRITICAL",
"investor_risk": "LOW | MEDIUM | HIGH",
"product_risk": "LOW | MEDIUM | HIGH",
"fraud_indicators": [],
"suitability_concerns": [],
"regulatory_authority": ["BCB", "CVM", "COAF"],
"regulatory_basis": [],
"finding": "",
"corrective_action": "",
"consumer_explanation": "",
"confidence": "LOW | MEDIUM | HIGH"
}
Task categories
| Category | Task | Examples |
|---|---|---|
| A | Financial risk assessment (crypto products) | 40 |
| B | Suitability analysis (investor profile vs product) | 30 |
| C | Fraud pattern detection | 30 |
| D | Regulator routing (BCB / CVM / COAF) | 20 |
| E | Consumer explanation (plain language) | 20 |
Training dataset
| Platform | Link |
|---|---|
| HuggingFace Dataset | Fernandosr85/adaption-brazil-crypto-regulatory-qa |
| Kaggle Dataset | finrisk-br-brazilian-crypto-investor-risk-dataset |
| Kaggle Notebook | FinRisk-BR |
Financial risk scenarios covered
| Scenario | Authority | Risk |
|---|---|---|
| Token com promessa de rendimento garantido | CVM · BCB | CRITICAL |
| Esquema de pirâmide com criptoativos | CVM · BCB · COAF | CRITICAL |
| Golpe via Pix para compra de criptoativos | BCB | CRITICAL |
| Coerção para transferência via Pix | BCB | CRITICAL |
| PSAV operando sem autorização BCB | BCB | HIGH |
| KYC insuficiente com risco de lavagem | BCB · COAF | HIGH |
| Tokenização de RWA sem registro CVM | CVM | HIGH |
| Produto de alta volatilidade para perfil conservador | CVM | HIGH |
| Stablecoin para câmbio não declarado | BCB | HIGH |
| Token de utilidade sem características de valor mobiliário | CVM | MEDIUM |
Regulatory coverage
| Authority | Role |
|---|---|
| BCB (Banco Central do Brasil) | PSAV authorization, AML/CFT, Pix fraud, foreign exchange |
| CVM (Comissão de Valores Mobiliários) | Suitability, securities tokens, public offerings |
| COAF | AML/CFT, PEP due diligence, suspicious activity |
| Federal | Lei 14.478/2022, Decreto 11.563/2023, Resolução Conjunta nº 6/2023 |
Credits
- Fine-tuning platform: Adaption — AutoScientist & Adaptive Data
- Challenge: AutoScientist Challenge 2026
- Training infrastructure: Adaption compute credits
- Author: Fernando Rodrigues · Kaggle: fernandosr85 · HuggingFace: Fernandosr85
Disclaimer
Experimental research artifact submitted to the AutoScientist Challenge 2026 (Finance category). Outputs do not constitute financial or legal advice and require review by qualified professionals.
- Task
- Text Generation
- Library
- peft
- License
- llama3.3
- Languages
- pt
- Created
- Jun 10, 2026
- Updated
- Jul 4, 2026