alinia/salamandra-7b-aligned-EADOP

Salamandra 7B aligned EADOP Model Card

Salamandra 7B aligned EADOP is a full-finetuning version of BSC Language Technologies Unit's Salamndra Instruct 7B model by the at the Barcelona Supercomputing Center focused on improving the handling of out-of-domain Questions in a RAG instruction-following setting.

The model has been finetuned on a dataset consisting of 2,000+ human annotated in- and out-of-domain user messages and assistant responses in the context of a chatbot that can provide helpful information about the current Catalan legislation. The dataset alinia/EADOP-RAG-out-of-domain was collected in collaboration with the Entitat Autònoma del Diari Oficial i de Publicacions (EADOP) and it consists of user messages and assistant responses in Catalan and Spanish.

DISCLAIMER: This model is a proof-of-concept designed to demonstrate the effects of finetuning an Instruction model with a small dataset of out-of-domain questions in the model's capability to politely and informatively refuse to answer questions that are out-of-domain. As a proof-of-concept, the model is still prone to generate harmful or inappropriate content.

---

Model Details

Please refer to the Salamndra Instruct 7B model details for the specific details about the model architecture and pretraining.

Intended Use

This model was developed as a proof-of-concept to demonstrate the effects of finetuning an Instruction model with a small dataset of in- and out-of-domain questions in the model's capability to politely and informatively refuse to answer questions that are out-of-domain in the context of a domain-specific RAG-based chatbot.

How to use

This model uses the ChatML, the same instruction-following conversation format as the base model.

from datetime import datetime
from transformers import AutoTokenizer, AutoModelForCausalLM
import transformers
import torch

model_id = "BSC-LT/salamandra-7b-instruct"

text = "At what temperature does water boil?"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype=torch.bfloat16
  )

message = [ { "role": "user", "content": text } ]

prompt = tokenizer.apply_chat_template(
    message,
    tokenize=False,
    add_generation_prompt=True
)

inputs = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt")
outputs = model.generate(input_ids=inputs.to(model.device), max_new_tokens=200)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Using this template, each turn is preceded by a <|im_start|> delimiter and the role of the entity (either user, for content supplied by the user, or assistant for LLM responses), and finished with the <|im_end|> token.

---

Finetuning Data

Please refer to alinia/EADOP-RAG-out-of-domain for the Dataset Card.

Author

This model has been finetuned by Alinia AI.

Contact

For further information, please email [email protected].

Acknowledgements

This project is part of a partnership with the Language Technologies Unit at the Barcelona Supercomputing Center. The data collection process was supported by the Entitat Autònoma del Diari Oficial i de Publicacions (EADOP).