AI-Scientist-v2/bfts_config.yaml at main · SakanaAI/AI-Scientist-v2 · 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
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
# path to the task data directory
data_dir: "data"
preprocess_data: False

goal: null
eval: null

log_dir: logs
workspace_dir: workspaces

# whether to copy the data to the workspace directory (otherwise it will be symlinked)
# copying is recommended to prevent the agent from accidentally modifying the original data
copy_data: True

exp_name: run # a random experiment name will be generated if not provided

# settings for code execution
exec:
  timeout: 3600
  agent_file_name: runfile.py
  format_tb_ipython: False

generate_report: True
# LLM settings for final report from journal
report:
  model: gpt-4o-2024-11-20
  temp: 1.0

experiment:
  num_syn_datasets: 1

debug:
  stage4: False

# agent hyperparams
agent:
  type: parallel
  num_workers: 4
  stages:
    stage1_max_iters: 20
    stage2_max_iters: 12
    stage3_max_iters: 12
    stage4_max_iters: 18
  # how many improvement iterations to run
  steps: 5 # if stage-specific max_iters are not provided, the agent will use this value for all stages
  # whether to instruct the agent to use CV (set to 1 to disable)
  k_fold_validation: 1
  multi_seed_eval:
    num_seeds: 3 # should be the same as num_workers if num_workers < 3. Otherwise, set it to be 3.
  # whether to instruct the agent to generate a prediction function
  expose_prediction: False
  # whether to provide the agent with a preview of the data
  data_preview: False

  # LLM settings for coding
  code:
    model: anthropic.claude-3-5-sonnet-20241022-v2:0
    temp: 1.0
    max_tokens: 12000

  # LLM settings for evaluating program output / tracebacks
  feedback:
    model: gpt-4o-2024-11-20
    # gpt-4o
    temp: 0.5
    max_tokens: 8192

  vlm_feedback:
    model: gpt-4o-2024-11-20
    temp: 0.5
    max_tokens: null

  search:
    max_debug_depth: 3
    debug_prob: 0.5
    num_drafts: 3

  # Options for summarizing findings and selecting the best node
  # If not specified, the default behavior will be used.

  # summary:
  #   model: gpt-4o
  #   temp: 0.3

  # select_node:
  #   model: gpt-4o
  #   temp: 0.3