license: apache-2.0
language:
- it
- en
library_name: transformers
tags:
- sft
- it
- mistral
- chatml
Model Information
XXXX is an updated version of Mistral-7B-v0.2, specifically fine-tuned with SFT and LoRA adjustments.
- It's trained both on publicly available datasets, like SQUAD-it, and datasets we've created in-house.
- it's designed to understand and maintain context, making it ideal for Retrieval Augmented Generation (RAG) tasks and applications requiring contextual awareness.
Evaluation
We evaluated the model using the same test sets as used for the Open Ita LLM Leaderboard
| hellaswag_it acc_norm | arc_it acc_norm | m_mmlu_it 5-shot acc | Average |
|---|---|---|---|
| 0.6067 | 0.4405 | 0.5112 | 0,52 |
Usage
Be sure to have transformers, peft and sentencepiece installed
pip install transformers peft sentencepiece
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel, PeftConfig
device = "cuda"
config = PeftConfig.from_pretrained("MoxoffSpA/xxxx")
model = AutoModelForCausalLM.from_pretrained("alpindale/Mistral-7B-v0.2-hf")
tokenizer = AutoTokenizer.from_pretrained("alpindale/Mistral-7B-v0.2-hf")
model = PeftModel.from_pretrained(model, "MoxoffSpA/xxxx")
messages = [
{"role": "user", "content": "Qual è il tuo piatto preferito??"},
{"role": "assistant", "content": "Beh, ho un debole per una buona porzione di risotto allo zafferano. È un piatto che si distingue per il suo sapore ricco e il suo bellissimo colore dorato, rendendolo irresistibile!"},
{"role": "user", "content": "Hai delle ricette con il risotto che consigli?"}
]
encodeds = tokenizer.apply_chat_template(messages, return_tensors="pt")
model_inputs = encodeds.to(device)
model.to(device)
generated_ids = model.generate(model_inputs, max_new_tokens=1000, do_sample=True)
decoded = tokenizer.batch_decode(generated_ids)
print(decoded[0])
Bias, Risks and Limitations
xxxx has not been aligned to human preferences for safety within the RLHF phase or deployed with in-the-loop filtering of responses like ChatGPT, so the model can produce problematic outputs (especially when prompted to do so). It is also unknown what the size and composition of the corpus was used to train the base model (mistralai/Mistral-7B-v0.2), however it is likely to have included a mix of Web data and technical sources like books and code.
Links to resources
- SQUAD-it dataset: https://huggingface.co/datasets/squad_it
- Mistral_7B_v0.2 original weights: https://models.mistralcdn.com/mistral-7b-v0-2/mistral-7B-v0.2.tar
- Mistral_7B_v0.2 model: https://huggingface.co/alpindale/Mistral-7B-v0.2-hf
- Open Ita LLM Leaderbord: https://huggingface.co/spaces/FinancialSupport/open_ita_llm_leaderboard
Quantized versions
We have published as well the 4 bit and 8 bit versions of this model: https://huggingface.co/MoxoffSpA/xxxxQuantized/main
The Moxoff Team
Jacopo Abate, Marco D'Ambra, Gianpaolo Francesco Trotta, Luigi Simeone