metadata
license: cc-by-nc-4.0
datasets:
- nickrosh/Evol-Instruct-Code-80k-v1
- MBZUAI/LaMini-instruction
language:
- en
base_model:
- Bertug1911/BrtGPT-1-Pre
pipeline_tag: text-generation
tags:
- code
BrtGPT-1-Pre-Code
Model Summary
We're introducing "BratGPT-1-Pre-Code"! Our model was retrained using the "BrtGPT-1-Pre" model, which was already pre-trained, using code data.
Compared to the BrtGPT-1-Pre model, it can write much better code, even with typos.
No change was observed in general/daily chat and simple knowledge-based question-and-answer capabilities.
It may produce some harmful output.
Difference Between Models
Examples:
| Prompt | BrtGPT-1-Pre |
|---|---|
| "Write me a code that prints "Hello World". | "Here's a code that prints "Hello World" in a list of words:```for i in range(1, 2, 3, 4, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5," |
| "Write me a code that generates random number." | Code: |
def random(1):
return random(1)
BrtGPT-1-Pre-Code's answers:
1- "Write me a code that prints "Hello World" Code:
```python
def print_hello_numbers(numbers):
if num < num:
return num
elif num % num % num % num % num % num % num % num % num % num % num % num % num % num % num % num % num
2- "Write me a code that generates random number. Code:
#Here is a code that generates random number in python 3:
```python
def generate_random_number(num):
# Create a new random number between 1 and 1
random_number = random.randint(num)
random_number = random.randint(num)
random_number = random.randint(num)
# Create a new
How to use?
You can run this code to use:
import torch
from transformers import PreTrainedTokenizerFast, GPT2LMHeadModel
def extract_response_between_tokens(text: str) -> str:
start_token = "<|im_start|>assistant<|im_sep|>"
end_token = "<|im_end|>"
try:
start_idx = text.index(start_token) + len(start_token)
end_idx = text.index(end_token, start_idx)
return text[start_idx:end_idx]
except ValueError:
# Tokenlar bulunamazsa orijinal metni döndür
return text
if __name__ == "__main__":
model_name_or_path = "Bertug1911/BrtGPT-1-Pre"
tokenizer = PreTrainedTokenizerFast.from_pretrained(model_name_or_path)
model = GPT2LMHeadModel.from_pretrained(model_name_or_path)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
model.eval()
user_input = input("Enter something to ask model: ")
messages = [{"role": "user", "content": user_input}]
formatted_prompt = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True
)
inputs = tokenizer(formatted_prompt, return_tensors="pt").to(device)
generated = inputs["input_ids"]
# Generate config
max_new_tokens = 128
do_sample = True
top_k = 40
temperature = 0.8
im_end_token_id = tokenizer.convert_tokens_to_ids("<|im_end|>")
with torch.no_grad():
for i in range(max_new_tokens):
outputs = model(generated)
logits = outputs.logits[:, -1, :]
logits = logits / temperature
if top_k > 0:
top_k_values, top_k_indices = torch.topk(logits, top_k)
logits_filtered = torch.full_like(logits, float('-inf'))
logits_filtered.scatter_(1, top_k_indices, top_k_values)
logits = logits_filtered
probs = torch.softmax(logits, dim=-1)
if do_sample:
next_token = torch.multinomial(probs, num_samples=1)
else:
next_token = torch.argmax(probs, dim=-1, keepdim=True)
generated = torch.cat([generated, next_token], dim=1)
if next_token.item() == im_end_token_id:
break
output = tokenizer.decode(generated[0], skip_special_tokens=False)
# Special token conversions
no_spaces = output.replace(" ", "")
step2 = no_spaces.replace("Ġ", " ")
formatted_output = step2.replace("Ċ", "\n")
if not formatted_output.strip().endswith("<|im_end|>"):
formatted_output += "<|im_end|>"
assistant_response = extract_response_between_tokens(formatted_output)
print("\nModel output:\n", assistant_response)
Evulation
Evulation results is cooming soon!
Risks and biases
Model may generates:
- Illegal outputs
- Harmfull contents
Use with caution!!