Introduction

This is a very early work-in-progress prototype of a LLM. The goal is to generate Java code to perform proof-of-reserve of a blockchain address.

Proof-of-reserve is used to verify ownership of a blockchain address. Otherwise said, the owner of the address has the private key to the address. In the proof-of-reserve process, a challenge is sent to the owner of the address. The owner then signs the challenge with the private key. The signature is verified with the public key.

The goal of the project is therefore to generate the Java code that will process the verification. The prompt is in natural language. We assume the user does not have technical knowledge on the signature algorithm.

Dataset

The prompt_to_signature_verifier_code_dataset.jsonl is only an example on how the dataset should be formatted.

Examples of usage

Launch server locally

python app.py --port 5000

Build and launch container

docker build -t blockchain-verifier-generator . docker run -p 5000:5000 blockchain-verifier-generator

Training

python main.py --train --dataset_path "../data/prompt_to_signature_verifier_code_dataset.jsonl" --epochs 3 --batch_size 1 --output_dir "../model/gpt2-java-trained"

curl -X POST "http://127.0.0.1:5000/train" -H "Content-Type: application/json"
-d '{"dataset_path": "../data/prompt_to_signature_verifier_code_dataset.jsonl", "output_dir": "../model/gpt2-java-trained", "epochs": 3, "batch_size": 1}'

Pruning

python main.py --prune --saved_model_path ../model/gpt2-java-trained --prune_method early_late --output_dir ../model/gpt2-java-pruned

curl -X POST "http://127.0.0.1:5000/prune" -H "Content-Type: application/json"
-d '{"saved_model_path": "../model/gpt2-java-trained", "prune_method": "early_late", "output_dir": "../model/gpt2-java-pruned"}'

Quantization

python main.py --quantize --saved_model_path ../model/gpt2-java-trained --quantization_method 8bit --output_dir ../model/gpt2-java-quantized

curl -X POST "http://127.0.0.1:5000/quantize" -H "Content-Type: application/json"
-d '{"saved_model_path": "../model/gpt2-java-trained", "quantization_method": "8bit", "output_dir": "../model/gpt2-java-quantized"}'

Generating

python main.py --generate --saved_model_path "../model/gpt2-java-trained" --prompt 'Generate a Solana verifier'

curl -X POST "http://127.0.0.1:5000/generate" -H "Content-Type: application/json"
-d '{"saved_model_path": "../model/gpt2-java-trained", "json_input": "'