A fine-tuned version of Meta Llama 3.1 8B Instruct specialized in Fhenix blockchain development and Fully Homomorphic Encryption (FHE) smart contracts.

Model Details

Property	Value
Base model	meta-llama/Llama-3.1-8B-Instruct
Fine-tuning method	LoRA (PEFT)
Quantization	4-bit (NF4)
Trainable parameters	41.9M / 8.07B (0.52%)
Training epochs	3
Final train loss	1.07 → 0.52
Final eval loss	0.601
Training time	~30 minutes on RTX 5060 Ti

Training Data

335 curated Fhenix Q&A pairs (268 train / 33 val / 34 test)
Topics: FHE.sol operations, @cofhe/sdk, encrypted types, Hardhat testing, frontend integration, access control, permits, gas optimization

Capabilities

FHE.sol arithmetic, comparison, bitwise, and conditional operations
Encrypted types: ebool, euint8/16/32/64/128, eaddress
@cofhe/sdk client library — encryption, decryption, permits
Smart contract development with encrypted storage
Hardhat testing and deployment for FHE contracts
Frontend integration with React/Next.js
Debugging and troubleshooting FHE applications
Gas optimization techniques specific to FHE

Usage

python

from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel
import torch

# Load base model in 4-bit
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
)
base_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B-Instruct",
    quantization_config=bnb_config,
    device_map="auto"
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "shree291/fhenix-llama3-lora")
tokenizer = AutoTokenizer.from_pretrained("shree291/fhenix-llama3-lora")
model.eval()

# Ask a question
messages = [
    {"role": "system", "content": "You are an expert Fhenix blockchain developer assistant specializing in Fully Homomorphic Encryption (FHE) smart contracts."},
    {"role": "user", "content": "How do I encrypt a uint32 value in a Solidity contract?"}
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=300, temperature=0.7, do_sample=True)

answer = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(answer)

Example Outputs

Q: What is FHE.select and when should I use it?

FHE.select is a conditional operator that selects between two encrypted values based on an encrypted boolean condition — the FHE equivalent of an if-else statement. Use it when you need conditional logic without revealing the condition. Example: euint32 result = FHE.select(isGreater, a, b);

Q: What is the role of the Threshold Network in Fhenix?

The Threshold Network provides decentralized threshold decryption. Multiple independent nodes collectively decrypt values using threshold cryptography — no single node can decrypt alone. It validates permits and ACL permissions before decrypting, secured via EigenLayer staking.

LoRA Configuration

json

{
  "r": 16,
  "lora_alpha": 32,
  "target_modules": ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
  "lora_dropout": 0.05,
  "bias": "none",
  "task_type": "CAUSAL_LM"
}

Limitations

Specialized for Fhenix/CoFHE — not a general-purpose coding assistant
Based on Llama 3.1 license terms — requires accepting Meta's license
Small dataset (335 pairs) — may hallucinate on edge cases not covered in training

License

This adapter follows the Llama 3.1 Community License.