File size: 2,287 Bytes
574c97f
 
 
 
298c26a
574c97f
298c26a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
574c97f
 
 
 
 
a6f080d
 
 
 
 
 
 
 
574c97f
 
298c26a
574c97f
 
 
 
 
 
 
298c26a
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
---
license: apache-2.0
---

# Kexer models

Kexer models is a collection of fine-tuned open-source generative text models fine-tuned on Kotlin Exercices dataset. 
This is a repository for fine-tuned CodeLlama-7b model in the Hugging Face Transformers format.

# Model use

```
  from transformers import AutoModelForCausalLM, AutoTokenizer

  # Load pre-trained model and tokenizer
  model_name = 'JetBrains/CodeLlama-7B-Kexer'  # Replace with the desired model name
  tokenizer = AutoTokenizer.from_pretrained(model_name)
  model = AutoModelForCausalLM.from_pretrained(model_name).cuda()

  # Encode input text
  input_text = """This function takes an integer n and returns factorial of a number:
  fun factorial(n: Int): Int {"""
  input_ids = tokenizer.encode(input_text, return_tensors='pt').to('cuda')

  # Generate text
  output = model.generate(input_ids, max_length=150, num_return_sequences=1, no_repeat_ngram_size=2, early_stopping=True)

  # Decode and print the generated text
  generated_text = tokenizer.decode(output[0], skip_special_tokens=True)
  print(generated_text)
```

# Training setup

The model was trained on one A100 GPU with following hyperparameters:

|         **Hyperparameter**           |             **Value**              |
|:---------------------------:|:----------------------------------------:|
|           `warmup`            |           10%            |
|        `max_lr`        |          1e-4          |
|        `scheduler`        |          linear          |
|        `total_batch_size`        |          256 (~130K tokens per step)          |


# Fine-tuning data

For this model we used 15K exmaples of Kotlin Exercices dataset {TODO: link!}. For more information about the dataset follow th link.

# Evaluation 

To evaluate we used Kotlin Humaneval (more infromation here)

Fine-tuned model:

|         **Model name**           |             **Kotlin HumanEval Pass Rate**              |             **Kotlin Completion**              |
|:---------------------------:|:----------------------------------------:|:----------------------------------------:|
|           `base model`            |           26.89            |           0.388            |
|        `fine-tuned model`        |          42.24         |          0.344          |