metadata

language:
  - it
license: cc-by-nc-4.0
tags:
  - sft
  - it
  - mistral
  - chatml
  - axolotl
prompt_template: >-
  <|im_start|>system {system_message}<|im_end|> <|im_start|>user
  {prompt}<|im_end|> <|im_start|>assistant
model-index:
  - name: maestrale-chat-v0.4-alpha-sft
    results: []

Want to contribute? Please donate! This will let us work on better datasets and models!

Maestrale chat alpha ༄

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, merged with occiglot.
Fine-Tuning: SFT performed on 1.7M convs/instructions for 2 epochs.

v0.4

Agent
Improved truthfullness
Improved Math & Reasoning capabilities
Mermaid mindmaps
More latin translations, poems, ...

This model uses ChatML prompt format:

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

Scores

Tasks	Version	Filter	n-shot	Metric	Value		Stderr
hellaswag_it	1	none	0	acc	0.5220	±	0.0052
		none	0	acc_norm	0.6887	±	0.0048
arc_it	1	none	0	acc	0.1762	±	0.0111
		none	0	acc_norm	0.5090	±	0.0146
m_mmlu_it	0	none	5	acc	0.569	±	0.0043

Usage:

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

tokenizer = AutoTokenizer.from_pretrained("mii-llm/maestrale-chat-v0.4-alpha-sft")
model = AutoModelForCausalLM.from_pretrained("mii-llm/maestrale-chat-v0.4-alpha-sft", 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|>")
)

streamer = TextStreamer(tokenizer, skip_prompt=True)

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

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

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

Intended uses & limitations

It's an alpha version; it's not safe, but it can refuse to answer.