Cabuxa-7B

Cabuxa is a LLaMA-7B model for Galician that can answer instructions in the Alpaca format.

It was fed with the 80% of the irlab-udc/alpaca_data_galician dataset, as we are keeping the remaining 20% for future evaluation and research.

This work broadens the Portuguese effort from 22h/cabrita-lora-v0-1 to Galician. Our working notes are available here.

How to Get Started with Cabuxa-7B

Use the code below to get started with the model.

from peft import PeftModel
from transformers import AutoModelForCausalLM, LlamaTokenizer, GenerationConfig

config = PeftConfig.from_pretrained("irlab-udc/cabuxa-7b")
model = AutoModelForCausalLM.from_pretrained("huggyllama/llama-7b", device_map="auto")
model = PeftModel.from_pretrained(model, "irlab-udc/cabuxa-7b")
tokenizer = LlamaTokenizer.from_pretrained("huggyllama/llama-7b")

# This function builds the prompt in Alpaca format
def generate_prompt(instruction, input=None):
    if input:
        return f"""Abaixo está unha instrución que describe unha tarefa, xunto cunha entrada que proporciona máis contexto. 
               Escribe unha resposta que responda adecuadamente a entrada.
               ### Instrución:
               {instruction}
               ### Entrada:
               {input}
               ### Resposta:"""
    else:
        return f"""Abaixo está unha instrución que describe unha tarefa.
               Escribe unha resposta que responda adecuadamente a entrada.
               ### Instrución:
               {instruction}
               ### Resposta:"""


def evaluate(instruction, input=None):
    prompt = generate_prompt(instruction, input)
    inputs = tokenizer(prompt, return_tensors="pt")
    input_ids = inputs["input_ids"].cuda()
    generation_output = model.generate(
        input_ids=input_ids,
        generation_config=GenerationConfig(do_sample=True),
        return_dict_in_generate=True,
        output_scores=True,
        max_new_tokens=256,
    )
    for s in generation_output.sequences:
        output = tokenizer.decode(s)
        print("Resposta:", output.split("### Resposta:")[1].strip())

evaluate("Cal é a fórmula química da auga?")
evaluate(
    "Convence ao lector por que é importante un determinado tema.",
    "Por que é esencial priorizar o sono?",
)

Resposta: A fórmula química da auga é H₂O.

Resposta: O sono é esencial para todos os humanos, pero tamén é unha ferramenta importante para lograr obxectivos, aumentar a productividade, maximizar os beneficios do soño e mantenerse saudable.

Training

Configurations and Hyperparameters

The following LoraConfig config was used during training:

• r: 8
• lora_alpha: 16
• target_modules: ["q_proj", "v_proj"]
• lora_dropout: 0.05
• bias: "none"
• task_type: "CAUSAL_LM"

The following TrainingArguments config was used during training:

• per_device_train_batch_size: 64
• gradient_accumulation_steps: 32
• warmup_steps: 100
• num_train_epochs: 20
• learning_rate: 3e-4
• fp16=True

The following bitsandbytes quantization config was used during training:

• quant_method: bitsandbytes
• load_in_8bit: True
• load_in_4bit: False
• llm_int8_threshold: 6.0
• llm_int8_skip_modules: None
• llm_int8_enable_fp32_cpu_offload: False
• llm_int8_has_fp16_weight: False
• bnb_4bit_quant_type: fp4
• bnb_4bit_use_double_quant: False
• bnb_4bit_compute_dtype: float32

Loss

Epoch	Loss
0.98	2.6109
1.97	2.0596
2.95	1.5092
3.93	1.379
4.92	1.2849
5.9	1.208
6.88	1.1508
7.86	1.117
8.85	1.0873
9.83	1.0666
10.81	1.0513
11.8	1.0365
12.78	1.0253
13.76	1.0169
14.75	1.0118
15.73	1.0035
16.71	0.9968
17.7	0.9983
18.68	0.9924
19.66	0.9908

Framework versions

• PyTorch 2.1.0
• PEFT 0.6.0.dev0
• 🤗 Transformers 4.34.0
• 🤗 Datasets 2.14.5
• 🤗 Tokenizers 0.14.0

Environmental Impact

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

• Hardware Type: NVIDIA RTX A6000.
• Hours used: 72.
• Cloud Provider: Private infrastructure.
• Carbon Emitted: 9.33 Kg. CO2 eq.