Zenos GPT-J 6B Instruct 4-bit

Model Overview

Name: zenos-gpt-j-6B-instruct-4bit
Datasets Used: Alpaca Spanish, Evol Instruct
Architecture: GPT-J
Model Size: 6 Billion parameters
Precision: 4 bits
Fine-tuning: This model was fine-tuned using Low-Rank Adaptation (LoRa).
Content Moderation: This model is not moderated.

Description

Zenos GPT-J 6B Instruct 4-bit is a Spanish Instruction capable model based on the GPT-J architecture with 6 billion parameters. It has been fine-tuned on the Alpaca Spanish and Evol Instruct datasets, making it particularly suitable for natural language understanding and generation tasks in Spanish.

An experimental Twitter (X) bot is available at https://twitter.com/ZenosBot which makes comments on news published in media outlets from Argentina.

Requirements

The latest development version of Transformers, which includes serialization of 4 bits models.

Transformers
Bitsandbytes >= 0.41.3

Since this is a compressed version (4 bits), it can fit into ~7GB of VRAM.

Usage

You can use this model for various natural language processing tasks such as text generation, summarization, and more. Below is an example of how to use it in Python with the Transformers library:

from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig

# Load the tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("webpolis/zenos-gpt-j-6B-instruct-4bit")
model = AutoModelForCausalLM.from_pretrained(
    "webpolis/zenos-gpt-j-6B-instruct-4bit",
    use_safetensors=True
)

user_msg = '''Escribe un poema breve utilizando los siguientes conceptos:

Bienestar, Corriente, Iluminación, Sed'''

# Generate text; watch out the padding between [INST] ... [/INST]
prompt = f'[INST] {user_msg} [/INST]'

inputs = tokenizer(prompt, return_tensors="pt")
input_ids = inputs["input_ids"].to(model.device)
attention_mask = inputs["attention_mask"].to(model.device)

generation_config = GenerationConfig(
    temperature=0.2,
    top_p=0.8,
    top_k=40,
    num_beams=1,
    repetition_penalty=1.3,
    do_sample=True
)

with torch.no_grad():
    generation_output = model.generate(
        input_ids=input_ids,
        pad_token_id=tokenizer.eos_token_id,
        attention_mask=attention_mask,
        generation_config=generation_config,
        return_dict_in_generate=True,
        output_scores=False,
        max_new_tokens=512,
        early_stopping=True
    )

s = generation_output.sequences[0]
output = tokenizer.decode(s)
start_txt = output.find('[/INST]') + len('[/INST]')
end_txt = output.find("<|endoftext|>", start_txt)
answer = output[start_txt:end_txt]

print(answer)

Inference

Online

Currently, the HuggingFace's Inference Tool UI doesn't properly load the model. However, you can use it with regular Python code as shown above once you meet the requirements.

CPU

Best performance can be achieved downloading the GGML 4 bits model and doing inference using the rustformers' llm tool.

Requirements

For optimal performance:

4 CPU cores
8GB RAM

In my Core i7 laptop it goes around 250ms per token:

Acknowledgments

This model was developed by Nicolás Iglesias using the Hugging Face Transformers library.

LICENSE

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this software except in compliance with the License. You may obtain a copy of the License at

Apache License 2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.