Uploaded model
- Developed by: Enzo Palmisano
- License: mit
- Finetuned from model : microsoft/Phi-3-mini-4k-instruct
Evaluation
For a detailed comparison of model performance, check out the Leaderboard for Italian Language Models.
Here's a breakdown of the performance metrics:
| Metric | hellaswag_it acc_norm | arc_it acc_norm | m_mmlu_it 5-shot acc | Average |
|---|---|---|---|---|
| Accuracy Normalized | 0.6088 | 0.4440 | 0.5667 | 0.5398 |
How to Use
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig
import torch
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
tokenizer = AutoTokenizer.from_pretrained("e-palmisano/Phi3-ITA-mini-4k-instruct", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("e-palmisano/Phi3-ITA-mini-4k-instruct", trust_remote_code=True)
model.to(device)
generation_config = GenerationConfig(
penalty_alpha=0.6, # The values balance the model confidence and the degeneration penalty in contrastive search decoding.
do_sample = True, # Whether or not to use sampling ; use greedy decoding otherwise.
top_k=5, # The number of highest probability vocabulary tokens to keep for top-k-filtering.
temperature=0.001, # The value used to modulate the next token probabilities.
repetition_penalty=1.7, # The parameter for repetition penalty. 1.0 means no penalty.
max_new_tokens = 64, # The maximum numbers of tokens to generate, ignoring the number of tokens in the prompt.
eos_token_id=tokenizer.eos_token_id, # The id of the *end-of-sequence* token.
pad_token_id=tokenizer.eos_token_id, # The id of the *padding* token.
)
def generate_answer(question):
messages = [
{"role": "user", "content": question},
]
model_inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(device)
outputs = model.generate(model_inputs, generation_config=generation_config)
result = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
return result
question = """Quale è la torre più famosa di Parigi?"""
answer = generate_answer(question)
print(answer)