Karan6124/llama3-8b-multi-dataset-sft

Llama-3 8B Multi-Dataset SFT (LoRA)

This model is a fine-tuned version of Llama-3-8B aligned using Supervised Fine-Tuning (SFT) on a merged dataset of Alpaca, Dolly 15k, and OpenAssistant.

🚀 Key Features

• Training Framework: Unsloth (LoRA)
• Datasets: Merged Instruction tuning (Alpaca, Dolly, OASST)
• Quantization: 4-bit (bitsandbytes)
• Monitoring: Tracked via Weights & Biases

🔗 Project Links

• Full Pipeline Code: GitHub Repository
• Developer: Karan Shelar

🛠️ Usage

You can load this model using the unsloth library for ultra-fast inference:

from unsloth import FastLanguageModel
import torch

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "Karan6124/llama3-8b-multi-dataset-sft",
    max_seq_length = 2048,
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)

# Test prompt
instruction = "Write a clear Python function to check if a string is a palindrome."
prompt = f"### Instruction:\n{instruction}\n\n### Response:\n"

inputs = tokenizer([prompt], return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])

📊 Training Details

• Rank (r): 16
• Alpha: 32
• Learning Rate: 2e-4
• Optim: adamw_8bit