| --- |
| base_model: meta-llama/Llama-4-Scout-17B-16E-Instruct |
| library_name: peft |
| pipeline_tag: text-generation |
| tags: |
| - finance |
| - defi |
| - wallet-risk |
| - autoscientist |
| - lora |
| - peft |
| - foresight-ai |
| --- |
| |
| # Foresight AI DeFi Wallet Risk Classifier |
|
|
| This repository contains the AutoScientist-trained LoRA adapter for **Foresight AI**, a DeFi wallet risk review system built for the Adaption Labs AutoScientist Challenge in the finance category. |
|
|
| The adapter was trained to classify whether a DeFi wallet behavior sample indicates a higher-risk state within a 14-day horizon. The expected completion is a binary label: |
|
|
| - `0` means lower-risk outcome |
| - `1` means higher-risk outcome |
|
|
| ## Training Run |
|
|
| - Fine-tune job ID: `351792b9-5320-432e-ac12-889388d6b47a` |
| - Training experiment ID: `3d6e6540-6138-4890-b097-c29deb8e580f` |
| - Base model: `meta-llama/Llama-4-Scout-17B-16E-Instruct` |
| - Training method: SFT |
| - Adapter type: LoRA |
| - Data format: chat |
| - Epochs: 3 |
| - LoRA rank: 64 |
| - LoRA alpha: 128 |
| - Learning rate: 0.0001 |
| - Scheduler: cosine |
| - Final exported eval loss: 1.142578125 |
|
|
| ## Dataset |
|
|
| The model was trained on the Foresight AI DeFi wallet risk instruction dataset. The hackathon dataset contains 1,000 instruction rows produced from retrospective wallet behavior samples. |
|
|
| Important: these labels are retrospective proxy labels used for hackathon model development. They should not be described as production-verified liquidation or drawdown outcomes. |
|
|
| ## Intended Use |
|
|
| This adapter is intended for research and demonstration of DeFi wallet risk classification. It can support a user-facing wallet review workflow where outputs are presented as decision-support signals rather than financial advice. |
|
|
| ## Limitations |
|
|
| - This is not a production financial risk model. |
| - The training labels are retrospective proxy labels, not independently verified future outcomes. |
| - The model should not be used to execute trades, move funds, or make automated liquidation or lending decisions. |
| - The model does not prove protocol exposure, health factor, leverage, or liquidation risk unless those facts are supported by external on-chain data. |
| - Outputs should be paired with Foresight AI's evidence-aware wallet detector, provider status, and data quality checks. |
|
|
| ## Adapter Files |
|
|
| This export contains PEFT/LoRA adapter weights and tokenizer/config files: |
|
|
| - `adapter_model.safetensors` |
| - `adapter_config.json` |
| - `tokenizer.json` |
| - `tokenizer_config.json` |
| - `special_tokens_map.json` |
| - `chat_template.jinja` |
| - `trainer_state.json` |
|
|
| ## License And Base Model Terms |
|
|
| Use of this adapter is subject to the license and acceptable-use terms of the base model and any Adaption Labs challenge requirements. |
|
|
