Maestrale chat beta ༄

By @efederici and @mferraretto

Model description

• Language Model: Mistral-7b for the Italian language, continued pre-training for Italian on a curated large-scale high-quality corpus.
• Fine-Tuning: SFT performed on convs/instructions for three epochs.

v0.3

• Function calling
• Reduced default system prompt to avoid wasting tokens (pre-alignment)

This model uses ChatML prompt format:

<|im_start|>system
Sei un assistente utile.<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

Usage:

from transformers import (
    AutoTokenizer, 
    AutoModelForCausalLM, 
    GenerationConfig,
    TextStreamer
)
import torch

tokenizer = AutoTokenizer.from_pretrained("mii-llm/maestrale-chat-v0.3-beta")
model = AutoModelForCausalLM.from_pretrained("mii-llm/maestrale-chat-v0.3-beta", load_in_8bit=True, device_map="auto")

gen = GenerationConfig(
    do_sample=True,
    temperature=0.7,
    repetition_penalty=1.2,
    top_k=50,
    top_p=0.95,
    max_new_tokens=500,
    pad_token_id=tokenizer.eos_token_id,
    eos_token_id=tokenizer.convert_tokens_to_ids("<|im_end|>")
)

messages = [
    {"role": "system", "content": "Sei un assistente utile."},
    {"role": "user", "content": "{prompt}"}
]

with torch.no_grad(), torch.backends.cuda.sdp_kernel(
    enable_flash=True, 
    enable_math=False,
    enable_mem_efficient=False
):
    temp = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = tokenizer(temp, return_tensors="pt").to("cuda")

    streamer = TextStreamer(tokenizer, skip_prompt=True)

    _ = model.generate(
        **inputs,
        streamer=streamer,
        generation_config=gen
    )

Intended uses & limitations

It's a beta sft version, but it's not aligned. It's a first test. We are working on alignment data and evals.