falcon-40b-code-alpaca

<br><br>

<p align="center"> <img src="https://github.com/jina-ai/finetuner/blob/main/docs/_static/finetuner-logo-ani.svg?raw=true" alt="Finetuner logo: Finetuner helps you to create experiments in order to improve embeddings on search tasks. It accompanies you to deliver the last mile of performance-tuning for neural search applications." width="150px"> </p> <p align="center"> <b>LLM Generation models trained by Jina AI, Finetuner team.</b>

This repo contains the full weights (16bit) for Falcon-40b fit on the Code Alpaca dataset.

Reproduction

This version of the weights was trained with the following hyperparameters:

• Epochs: 2
• Batch size: 128
• Micro batch size: 4
• Learning rate: 3e-4
• Lora r: 8
• Lora target modules: query_key_value

You can reproduce using this repository:

https://github.com/jina-ai/jerboa

Make sure you install requirements and finetune using this command using the following command:

python finetune.py \
--base-model tiiuae/falcon-40b --lora-target-modules query_key_value \
--data-path sahil2801/CodeAlpaca-20k --output-dir ./lora-alpaca-code \
--batch-size 128 --micro-batch-size 4 --eval-limit 45 \
--eval-file code_eval.jsonl --wandb-project jerboa --wandb-log-model \
--wandb-watch gradients --num-epochs 2

Inference

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM


TOKENIZER_SOURCE = 'tiiuae/falcon-40b'
BASE_MODEL = 'jinaai/falcon-40b-code-alpaca'
DEVICE = "cuda"

PROMPT = """
Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
Write a for loop in python

### Input:

### Response:
"""
model = AutoModelForCausalLM.from_pretrained(
    pretrained_model_name_or_path=BASE_MODEL,
    torch_dtype=torch.float16,
    trust_remote_code=True,
    device_map='auto',
)

model.eval()

tokenizer = AutoTokenizer.from_pretrained(
    TOKENIZER_SOURCE,
    trust_remote_code=True,
    padding_side='left',
)
tokenizer.pad_token = tokenizer.eos_token

inputs = tokenizer(PROMPT, return_tensors="pt")
input_ids = inputs["input_ids"].to(DEVICE)
input_attention_mask = inputs["attention_mask"].to(DEVICE)

with torch.no_grad():
    generation_output = model.generate(
        input_ids=input_ids,
        attention_mask=input_attention_mask,
        return_dict_in_generate=True,
        max_new_tokens=32,
        eos_token_id=tokenizer.eos_token_id,
    )
generation_output = generation_output.sequences[0]
output = tokenizer.decode(generation_output, skip_special_tokens=True)

print(output)

Contact

Join our Discord community and chat with other community members about ideas.