MultiAgentTrainer/config.example.yaml at main · sroomberg/MultiAgentTrainer · 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
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
# MultiAgentTrainer configuration
# Copy to multiagenttrainer.yaml in your working directory.

# Named execution targets — referenced by training and finetuner.targets.
# Each machine can carry its own agent_command to right-size the model to
# the instance (e.g. run a large model on a big GPU host, a small model on
# a cheaper one).  Omit this section to run everything on the local machine.
machines:
  - name: gpu-large
    execution:
      type: ssh
      ssh_host: trainer1.example.com
      ssh_key: ~/.ssh/id_ed25519
      remote_dir: /tmp/mat-runs
    # Override the global agent_command to use a larger model on this host.
    agent_command: >-
      claude -p {prompt}
      --allowedTools "Bash,Read,Edit"
      --permission-mode acceptEdits
      --model claude-opus-4-7

  - name: gpu-small
    execution:
      type: ssh
      ssh_host: trainer2.example.com
      remote_dir: /tmp/mat-runs
    agent_command: >-
      claude -p {prompt}
      --allowedTools "Bash,Read,Edit"
      --permission-mode acceptEdits
      --model claude-haiku-4-5

autoresearch:
  repo: "https://github.com/karpathy/autoresearch"  # or a local path
  branch: "master"
  train_time: 300       # seconds per experiment (default 5 min)
  # program_md: ./my-program.md   # optional override for agent instructions

sources:
  # Local git repository
  - type: local_repo
    path: /home/user/my-project
    include: ["**/*.py", "**/*.md"]
    exclude: ["**/node_modules/**"]

  # Any git-cloneable URL
  - type: remote_repo
    url: "https://github.com/user/repo.git"
    branch: main

  # GitHub repository (web URL)
  - type: github_repo
    url: "https://github.com/user/repo"

  # All repos in a GitHub organisation
  - type: github_org
    url: "https://github.com/my-org"
    max_repos: 50
    visibility: all   # all | public | private

  # AWS Bedrock knowledge base
  - type: bedrock_knowledge_base
    knowledge_base_id: "ABCDEF1234"
    region: "us-east-1"
    query: "training data for code generation"
    max_results: 100

training:
  agent_command: 'claude -p {prompt} --allowedTools "Bash,Read,Edit" --permission-mode acceptEdits'
  max_experiments: 50
  output_dir: ./training-runs

# Failure notifications (optional)
# Requires: boto3 and an SES-verified sender address in your AWS account.
notifications:
  ses:
    from_email: alerts@example.com
    to_emails:
      - oncall@example.com
    region: us-east-1
    subject_prefix: "[MultiAgentTrainer]"

# Fine-tuning configuration (optional)
# Run: mat finetune start
finetuner:
  backend: opensource   # opensource | bedrock
  jobs_dir: ./finetune-jobs

  # Open-source: HuggingFace + PEFT/LoRA
  # Requires: pip install 'multiagenttrainer[opensource]'
  opensource:
    model_id: meta-llama/Llama-3.2-1B
    output_dir: ./finetuned-models
    lora_r: 16
    lora_alpha: 32
    lora_dropout: 0.05
    max_seq_length: 2048
    num_epochs: 3
    batch_size: 4
    learning_rate: 2e-4
    gradient_accumulation_steps: 4
    use_4bit: true                        # QLoRA (requires bitsandbytes + CUDA)
    target_modules: [q_proj, v_proj]      # or omit to use "all-linear"
    packing: true                         # bin chunks together to eliminate padding waste
    use_bf16: false                       # bf16 precision — recommended on Ampere+ (A100, RTX 30xx+)
    use_flash_attention: false            # Flash Attention 2 — requires: pip install flash-attn

  # AWS Bedrock model customization
  bedrock:
    base_model_id: amazon.titan-text-lite-v1
    region: us-east-1
    role_arn: arn:aws:iam::123456789012:role/BedrockFineTuningRole
    output_s3_uri: s3://my-bucket/finetuned-models/
    training_data_s3_uri: s3://my-bucket/training-data/
    customization_type: CONTINUED_PRE_TRAINING  # CONTINUED_PRE_TRAINING | FINE_TUNING
    epochs: 1
    batch_size: 8
    learning_rate: 1e-5
    job_name_prefix: mat-finetune

  # Multi-target fine-tuning: one job per (model, machine) pairing.
  # All targets start concurrently. Each target inherits the backend defaults
  # above and overrides only the fields listed here.
  # The machine field is metadata that documents which host this target is
  # sized for; assign each OpenSource target to a distinct GPU to avoid OOM.
  targets:
    - name: llama-3b-large
      model_id: meta-llama/Llama-3.2-3B
      machine: gpu-large        # big GPU for the bigger model
      backend: opensource
      num_epochs: 3
      batch_size: 2

    - name: llama-1b-small
      model_id: meta-llama/Llama-3.2-1B
      machine: gpu-small        # smaller GPU for the smaller model
      backend: opensource
      num_epochs: 5
      batch_size: 4

    - name: titan-bedrock
      model_id: amazon.titan-text-lite-v1
      backend: bedrock
      customization_type: FINE_TUNING