yuvrajpant56/Llama-3.1-8B-ultrachat-sft

LLaMA‑3.1 8B Instruct (Ultrachat SFT)

Fine-tuned on: Ultrachat 200k (train_sft)
Model size: 8 B parameters
Format: [safetensors]

🚀 Model Description

This is LLaMA‑3.1 8 B, instruction‑tuned on the Ultrachat dataset for more human‑like chat. It excels at back‑and‑forth dialogues, follows system/user/assistant markers, and has been lightly profanity‑filtered.

Use cases:

• Chatbots
• Instruction following
• Dialogue generation

📦 Technical Specifications

• Architecture: LLaMA‑3 (8 B)
• Rope scaling: dynamic, factor=2.0
• Precision: bf16
• Hardware used for SFT: 8×A100 GPUs
• Training epochs: 3
• Batch size: 1 × 16 accumulated
• Learning rate: 2 × 10⁻⁵ (cosine schedule, 100 warmup steps)

🔧 Usage

from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

tokenizer = AutoTokenizer.from_pretrained("yuvrajpant56/llama3-ultrachat-sft")
model     = AutoModelForCausalLM.from_pretrained("yuvrajpant56/llama3-ultrachat-sft", device_map="auto")

chat = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    max_new_tokens=128,
    do_sample=True,
    temperature=0.7,
)

prompt = "<|begin_of_sentence|>\nUser: Hello, how are you?\n<|end_of_sentence|>"
print(chat(prompt))