RegTech BR — Brazilian Crypto Legal Adapter
LoRA adapter for Brazilian crypto asset compliance reasoning, fine-tuned on Mixtral-8x7B-Instruct-v0.1 via Adaption's AutoScientist platform.
Covers regulatory scenarios under BCB · CVM · COAF and federal legislation (Lei 14.478/2022, Decreto 11.563/2023).
Evaluation results
Training Winrates
The adapted model outperforms the base model with 63% win rate vs 37% on held-out legal compliance evaluation — a +70% relative improvement over the base model.
| Model | Win Rate |
|---|---|
Base (Mixtral-8x7B-Instruct-v0.1) |
37% |
Adapted (brazilian_crypto_compliance) |
63% |
Train/Eval Metrics
| Metric | Value |
|---|---|
| Initial train loss | 1.601 |
| Final validation loss | 0.747 |
| Loss reduction | −53% |
| Peak learning rate | 1.65e-4 |
| Training steps | 69 |
| Eval checkpoints | 5 |
| LR scheduler | cosine (warmup 0.05) |
| Gradient norm (peak → stable) | 5.33e6 → ~5.57e5 |
Train loss (cyan) converged from 1.601 to 0.747 over 69 steps across 3 epochs. Validation loss (black dots) tracked closely, confirming generalization without overfitting. Learning rate followed cosine schedule with warmup, peaking at step ~10 then decaying to near-zero. Gradient norm stabilized after step 10, indicating stable optimization throughout.
Dataset quality
| Metric | Value |
|---|---|
| Dataset grade | A |
| Quality improvement | +6.7% (Adaption Adaptive Data remastering) |
| Total examples | 19 instruction/response pairs |
Model details
| Field | Value |
|---|---|
| Base model | mistralai/Mixtral-8x7B-Instruct-v0.1 (46.7B) |
| Trained model name | adaption_brazilian_crypto_compliance |
| Training method | Supervised Fine-Tuning (SFT) + LoRA |
| LoRA rank (r) | 32 |
| LoRA alpha | 64 |
| LoRA dropout | 0.05 |
| Trainable modules | all-linear |
| Epochs | 3 |
| Batch size | max |
| Learning rate | 1.5e-4 (cosine scheduler, 0.5 cycles) |
| Warmup ratio | 0.05 |
| Weight decay | 0.01 |
| Max grad norm | 1.0 |
| Train on inputs | false |
| Data format | Chat (instruction/response) |
Training dataset
Fernandosr85/adaption-brazilian-crypto-compliance
19 instruction/response pairs covering:
| Scenario | Authority | Risk |
|---|---|---|
| PSAV without BCB authorization | BCB | HIGH |
| No asset segregation | BCB | HIGH |
| KYC only above R$100k | BCB + COAF | HIGH |
| No AML/CFT officer | BCB + COAF | HIGH |
| USDT FX reporting omitted | BCB | HIGH |
| Revenue/voting token public offering | CVM | HIGH |
| RWA tokenization without CVM registration | CVM | HIGH |
| PEP without enhanced due diligence | COAF + BCB | HIGH |
| Fully compliant PSAV | BCB + COAF | LOW |
| Lei 14.478/2022 framework | Federal | LOW |
| Utility token (non-security) | CVM | REQUIRES_REVIEW |
Regulatory coverage
| Authority | Role |
|---|---|
| BCB (Banco Central do Brasil) | PSAV authorization, segregation, AML/CFT |
| CVM (Comissão de Valores Mobiliários) | Token securities classification |
| COAF | AML/CFT, PEP due diligence |
| Federal | Lei 14.478/2022, Decreto 11.563/2023 |
Model repositories
| Platform | Link |
|---|---|
| HuggingFace Model | Fernandosr85/regtech-br-legal-adapter |
| Kaggle Model | fernandosr85/regtech-br-brazilian-crypto-legal-adapter |
| HuggingFace Dataset | Fernandosr85/adaption-brazilian-crypto-compliance |
Demo
- HuggingFace Space: Fernandosr85/regtech-br
- Kaggle notebook: RegTech BR
- Kaggle dataset (RAG corpus): RegTech BR RAG Index
Credits
- Fine-tuning platform: Adaption — AutoScientist & Adaptive Data
- Challenge: AutoScientist Challenge 2026
- Training infrastructure: Adaption compute credits
- Dataset remastering: Adaption Adaptive Data pipeline (Grade A, +6.7% quality improvement)
- Author: Fernando Rodrigues · Kaggle: fernandosr85 · HuggingFace: Fernandosr85
Disclaimer
Experimental research artifact submitted to the AutoScientist Challenge 2026 (Legal category). Outputs do not constitute legal advice and require review by qualified professionals.
- Downloads last month
- -
Model tree for Fernandosr85/regtech-br-legal-adapter
Base model
mistralai/Mixtral-8x7B-v0.1