PatchLM/config.yaml at main · SmartSecLab/PatchLM · GitHub

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
# base_model: Salesforce/codet5-base
base_model: codellama/CodeLlama-7b-Instruct-hf
# base_model: codellama/CodeLlama-34b-Instruct-hf
# base_model: models/CodeLLama-7b-quantized-4bit
# most lightweight model of CodeLlama for instruction prompt
debug_mode: False
use_4bit_quantization: False
dataset_use: repairllama # FixMe or repairllama
only_compare: True
# compare with already fine-tune instruct_model
instruct_model: models/codellama20epoch-20250516-75358-full
# instruct_model: models/codellama-Debug
#
generation:
  max_new_tokens: 512
  do_sample: False
  temperature: 2.0
  num_beams: 10
  tokenizer: roberta # or auto
#
fine_tuning:
  output_dir: models
  learning_rate: 1e-4
  num_train_epochs: 20
  batch_size: 32 # 32 for eX3 HPC
  per_device_train_bsize: 8 # 16 for eX3 HPC
  weight_decay: 0.01
  logging_steps: 10
  max_steps: 10
# set the following paras if you are using FixMe dataset
preprocess:
  db_file: data/FixMe-v1.db
  max_hunks_per_url: 1
  # CodeT5 supported: Python, Java, JavaScript, PHP, Ruby, Go, C, and C#
  prog_lang: # list of programming languages
    - C
    - C++
    - Python
    - Java
    - JavaScript
    # - PHP
    # - Ruby
    # - Go
    # - C#