crucible/crucible.yaml at main · eren23/crucible · 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
# ─────────────────────────────────────────────────────────────────────────────
#  This is the Parameter Golf reference project config — the one Crucible
#  was originally built for. It is checked in as a working, end-to-end
#  example of an LM training setup on RunPod.
#
#  If you are starting a new project and want a clean, modality-agnostic
#  starting point instead, use:
#
#      cp crucible.yaml.example crucible.yaml
#      crucible project new my-project --template <lm|vision|diffusion|world_model|generic>
#
#  The `willdepueoai/parameter-golf` dataset reference below is a specific
#  HuggingFace dataset for the OpenAI Parameter Golf competition, NOT a
#  general default. Override `data.repo_id` for any other project.
# ─────────────────────────────────────────────────────────────────────────────
name: parameter-golf
version: "0.1.0"

# Compute provider
provider:
  type: runpod
  ssh_key: ~/.ssh/id_ed25519_runpod
  image: runpod/pytorch:1.0.2-cu1281-torch280-ubuntu2404
  gpu_types:
    - "NVIDIA GeForce RTX 4090"
  gpu_count: 1
  defaults:
    container_disk: 20
    volume_disk: 40
    workspace_path: /workspace/parameter-golf
    python_bin: python3.12
    env_source: .env.runpod.local

# Data pipeline
data:
  source: huggingface
  repo_id: willdepueoai/parameter-golf
  variant: fineweb10B_sp1024
  local_root: ./data
  manifest: manifest.json
  # Remote data readiness probe. Fleet bootstrap generates a probe
  # script from these fields; if any of the listed paths is missing,
  # `download_command` is run on the pod.
  probe:
    paths:
      - "data/tokenizers/fineweb_1024_bpe.model"
      - "data/datasets/fineweb10B_sp1024/fineweb_train_*.bin"
      - "data/datasets/fineweb10B_sp1024/fineweb_val_*.bin"
    download_command: "python3 data/cached_challenge_fineweb.py --variant sp1024"

# Training backends
training:
  - backend: torch
    script: src/crucible/training/torch_backend.py

# Metrics configuration
metrics:
  primary: val_bpb
  secondary: val_loss
  size: model_bytes
  direction: minimize

# Experiment presets
presets:
  smoke:
    MAX_WALLCLOCK_SECONDS: "180"
    ITERATIONS: "400"
    TRAIN_BATCH_TOKENS: "8192"
  screen:
    MAX_WALLCLOCK_SECONDS: "3600"
    ITERATIONS: "2000"
    TRAIN_BATCH_TOKENS: "65536"
    VAL_LOSS_EVERY: "500"
    VAL_BATCH_SIZE: "8192"
    TRAIN_LOG_EVERY: "20"
    WARMUP_STEPS: "10"
    TRAIN_SHARD_LIMIT: "8"
  proxy:
    MAX_WALLCLOCK_SECONDS: "1800"
    ITERATIONS: "6000"
    TRAIN_BATCH_TOKENS: "65536"

# Runner script on pods (invoked by fleet scheduler via SSH)
runner_script: src/crucible/runner/run_remote.py
remote_results_file: experiments.jsonl

# Timeout map (seconds) per backend per preset
timeout_map:
  torch:
    smoke: 300
    screen: 4200
    proxy: 2400
    medium: 4800
    promotion: 9600
    overnight: 5400
  mlx:
    smoke: 180
    screen: 900
    proxy: 3600
    medium: 7200
    promotion: 14400
    overnight: 7200

# Autonomous researcher
researcher:
  model: claude-sonnet-4-6-20250514
  budget_hours: 10.0
  max_iterations: 20
  program_file: program.md

# Version store
store_dir: .crucible

# Sync exclusions
sync_excludes:
  - .git
  - .venv
  - __pycache__
  - .crucible/designs
  - .crucible/context
  - .crucible/notes
  - .crucible/store.jsonl
  - .crucible/notes.jsonl
  - logs
  - data/datasets
  - data/tokenizers
  - fleet_runs
  - day_runs
  - .DS_Store
  - checkpoints
  - videos
  - wandb

# Output paths
results_file: experiments.jsonl
fleet_results_file: experiments_fleet.jsonl
logs_dir: logs
nodes_file: nodes.json