Kancil-V1-llama3-4bit

<!DOCTYPE html> <html lang="en"> <head> <meta charset="UTF-8"> <meta name="viewport" content="width=device-width, initial-scale=1.0"> <title>Document Title</title> <style> h1 { font-size: 36px; color: navy; font-family: 'Tahoma'; text-align: center; } </style> </head> <body> <h1>Introducing the Kancil family of open models</h1> </body> </html> <center> <img src="https://imgur.com/9nG5J1T.png" alt="Kancil" width="600" height="300"> <p><em>Kancil is a fine-tuned version of Llama 3 8B using synthetic QA dataset generated with Llama 3 70B. Version zero of Kancil is the first generative Indonesian LLM gain functional instruction performance using solely synthetic data.</em></p> <p><strong><a href="https://colab.research.google.com/drive/1Gp-I9vMqfhU_i5xX77ZKlh7eQCx1_13I?usp=sharing" style="color: blue; font-family: Tahoma;">❕Go straight to the colab demo❕</a></strong></p> <p><em style="color: black; font-weight: bold;">Beta preview</em></p> </center> Selamat datang!

I am ultra-overjoyed to introduce you... the 🦌 Kancil! It's a fine-tuned version of Llama 3 8B with the Tumpeng, an instruction dataset of 14.8 million words. Both the model and dataset is openly available in Huggingface.

📚 The dataset was synthetically generated from Llama 3 70B. A big problem with existing Indonesian instruction dataset is they're in reality not-very-good-translations of English datasets. Llama 3 70B can generate fluent Indonesian! (with minor caveats 😔)

🦚 This follows previous efforts for collection of open, fine-tuned Indonesian models, like Merak and Cendol. However, Kancil solely leverages synthetic data in a very creative way, which makes it a very unique contribution!

Version 1.0

This is the second working prototype, Kancil V1. ✨ Training

• 2.2x Dataset word count
• 2x lora parameters
• Rank-stabilized lora
• 2x fun

✨ New features

• Multi-turn conversation (beta; optimized for curhat/personal advice 😂)
• Better text generation (full or outline writing; optimized for essays)
• QA from text (copy paste to prompt and ask a question about it)
• Making slogans

This model was fine-tuned with QLoRA using the amazing Unsloth framework! It was built on top of unsloth/llama-3-8b-bnb-4bit and subsequently merged with the adapter.

Uses

This model is developed with research purposes for researchers or general AI hobbyists. However, it has one big application: You can have lots of fun with it!

Out-of-Scope Use

This is a research preview model with minimal safety curation. Do not use this model for commercial or practical applications.

You are also not allowed to use this model without having fun.

Getting started

As mentioned, this model was trained with Unsloth. Please use its code for better experience.

# Install dependencies. You need GPU to run this (at least T4)
%%capture
!pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
!pip install --no-deps xformers trl peft accelerate bitsandbytes
from unsloth import FastLanguageModel
import torch

# Available versions
KancilV1 = "catinthebag/Kancil-V1-llama3-4bit"

# Load the model
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = KancilV1,
    max_seq_length = 4096,
    dtype = None, # Auto detect
    load_in_4bit = True,
)

# This model was trained on this specific prompt template. Changing it might lead to performance degradations.
prompt_template = """<|user|>
{prompt}

<|assistant|>
{response}"""

# Start generating!
inputs = tokenizer(
[
prompt_template.format(
        prompt="Bagaimana cara memberi tahu orang tua kalau saya ditolak universitas favorit saya?",
        response="",)
], return_tensors = "pt").to("cuda")

outputs = model.generate(**inputs, max_new_tokens = 600, temperature=.8, use_cache = True)
print(tokenizer.batch_decode(outputs)[0].replace('\\n', '\n'))

Note: There is an issue with the dataset where the newline characters are interpreted as literal strings. Very sorry about this! 😔 Please keep the .replace() method to fix newline errors.

Acknowledgments

• Developed by: Afrizal Hasbi Azizy
• License: Llama 3 Community License Agreement