CRIA v1.3

💡 Article | 💻 Github | 📔 Colab 1,2

What is CRIA?

krē-ə plural crias. : a baby llama, alpaca, vicuña, or guanaco.

or what ChatGPT suggests, "Crafting a Rapid prototype of an Intelligent llm App using open source resources".

The initial objective of the CRIA project is to develop a comprehensive end-to-end chatbot system, starting from the instruction-tuning of a large language model and extending to its deployment on the web using frameworks such as Next.js.

Specifically, we have fine-tuned the llama-2-7b-chat-hf model with QLoRA (4-bit precision) using the mlabonne/CodeLlama-2-20k dataset. This fine-tuned model serves as the backbone for the CRIA chat platform.

📦 Model Release

CRIA v1.3 comes with several variants.

davzoku/cria-llama2-7b-v1.3: Merged Model
davzoku/cria-llama2-7b-v1.3-GGML: Quantized Merged Model
davzoku/cria-llama2-7b-v1.3_peft: PEFT adapter
davzoku/cria-llama2-7b-v1.3-GGUF: GGUF Format
- converted from ggml vq4_0 using python3 convert-llama-ggml-to-gguf.py -i ../text-generation-webui/models/cria/cria-llama2-7b-v1.3.ggmlv3.q4_0.bin -o cria-llama2-7b-v1.3.gguf
davzoku/cria-llama2-7b-v1.3-mlx: MLX Format in FP16
- Converted from davzoku/cria-llama2-7b-v1.3 using mlx-examples. see NOTES.md
davzoku/cria-llama2-7b-v1.3-q4-mlx: MLX Format in Q4
- Converted from davzoku/cria-llama2-7b-v1.3 using mlx-examples. see NOTES.md

🔧 Training

It was trained on a Google Colab notebook with a T4 GPU and high RAM.

Training procedure

The following bitsandbytes quantization config was used during training:

load_in_8bit: False
load_in_4bit: True
llm_int8_threshold: 6.0
llm_int8_skip_modules: None
llm_int8_enable_fp32_cpu_offload: False
llm_int8_has_fp16_weight: False
bnb_4bit_quant_type: nf4
bnb_4bit_use_double_quant: False
bnb_4bit_compute_dtype: float16

Framework versions

PEFT 0.4.0

💻 Usage

This model was converted to MLX format from davzoku/cria-llama2-7b-v1.3.

Use with mlx

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("davzoku/cria-llama2-7b-v1.3-mlx")
response = generate(model, tokenizer, prompt="hello", verbose=True)

Original Usage

# pip install transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "davzoku/cria-llama2-7b-v1.3"
prompt = "What is a cria?"

tokenizer = AutoTokenizer.from_pretrained(model)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    torch_dtype=torch.float16,
    device_map="auto",
)

sequences = pipeline(
    f'<s>[INST] {prompt} [/INST]',
    do_sample=True,
    top_k=10,
    num_return_sequences=1,
    eos_token_id=tokenizer.eos_token_id,
    max_length=200,
)
for seq in sequences:
    print(f"Result: {seq['generated_text']}")

References

We'd like to thank:

mlabonne for his article and resources on implementation of instruction tuning
TheBloke for his script for LLM quantization.

davzoku
/

cria-llama2-7b-v1.3-mlx