crucible/crucible.yaml.example 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
139
140
141
142
143
# ─────────────────────────────────────────────────────────────────────────────
#  Crucible project configuration — example template
#
#  Copy this file to `crucible.yaml` and edit it for your project, or run
#  `crucible init` to generate one automatically. This template is
#  modality-agnostic — works for LM, vision, diffusion, world models, etc.
#
#  To create a project spec for fleet runs, use:
#      crucible project new my-project --template <lm|vision|diffusion|world_model|generic>
#
#  For local-only development you don't need a project spec; this
#  `crucible.yaml` file is enough.
# ─────────────────────────────────────────────────────────────────────────────

name: my-project
version: "0.1.0"

# ─── Compute provider ───────────────────────────────────────────────────────
provider:
  type: runpod                           # "runpod" | "ssh"
  ssh_key: ~/.ssh/id_ed25519_runpod
  # image: runpod/pytorch:2.6.0-py3.11-cuda12.4.1-devel-ubuntu22.04
  # gpu_types: ["NVIDIA GeForce RTX 4090"]
  # gpu_count: 1
  # interruptible: true                  # false = secure/on-demand pods
  defaults:
    container_disk: 30
    volume_disk: 0
    workspace_path: /workspace/project
    python_bin: python3
    env_source: .env.runpod.local        # env file synced to pods

# ─── Data pipeline ──────────────────────────────────────────────────────────
# Configure for your dataset. Leave `repo_id` empty if you ship data via
# local_files on project specs instead.
data:
  source: huggingface                    # "huggingface" | "local_files" | "wandb_artifact"
  repo_id: ""                            # e.g. "HuggingFaceFW/fineweb-edu"
  local_root: ./data
  manifest: manifest.json
  # Remote data readiness probe. If `probe.paths` is empty (default),
  # fleet bootstrap skips the data step entirely. Populate it for
  # projects that ship their own dataset.
  probe:
    paths: []                            # e.g. ["data/train.h5", "data/val.h5"]
    # script: ""                         # custom readiness script (optional)
    download_command: ""                 # run when probe reports missing data

# ─── Fleet / SSH reliability ────────────────────────────────────────────────
# Per-step timeouts and initial-connect backoff. Defaults work for
# typical RunPod pods; tune step_timeouts["pip_install"] up if you
# have a big requirements.txt, or initial_connect.max_wait up for
# slow cloud types.
fleet:
  ssh:
    initial_connect:
      max_attempts: 6
      backoff_base: 5                    # seconds; doubles per attempt
      max_wait: 180                      # total budget
    step_timeouts:
      default: 300
      sync_repo: 600
      pip_install: 900
      data_download: 1800
      data_probe: 30

# ─── Training backends ──────────────────────────────────────────────────────
# Use `torch` for LM-style training, `generic` for diffusion / world models /
# custom trainers. Multiple backends can coexist.
training:
  - backend: torch
    script: train.py

# ─── Metrics ────────────────────────────────────────────────────────────────
metrics:
  primary: val_loss                      # metric the leaderboard ranks by
  direction: minimize                    # "minimize" | "maximize"
  # secondary: val_acc                   # optional display metric
  # size: model_bytes                    # for Pareto frontier analysis

# ─── Experiment presets ─────────────────────────────────────────────────────
# Presets are keyed by env var name — Crucible exports them before launching
# your training script. Add / rename to suit your workflow.
presets:
  smoke:                                 # 60s sanity check
    MAX_WALLCLOCK_SECONDS: "60"
    ITERATIONS: "400"
  screen:                                # 1h directional signal
    MAX_WALLCLOCK_SECONDS: "3600"
    ITERATIONS: "2000"
  proxy:                                 # 30min medium confidence
    MAX_WALLCLOCK_SECONDS: "1800"
    ITERATIONS: "6000"

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

# ─── Weights & Biases ───────────────────────────────────────────────────────
wandb:
  required: false                        # set true if every run must log to W&B
  project: ""                            # or set WANDB_PROJECT in env
  entity: ""                             # optional team/entity
  mode: online                           # "online" | "offline" | "disabled"

# ─── Execution policy ───────────────────────────────────────────────────────
execution_policy:
  require_remote: false                  # true = refuse to run on local machine
  required_provider: ""                  # e.g. "runpod"
  allow_local_dev: true

# ─── Version store ──────────────────────────────────────────────────────────
store_dir: .crucible

# ─── Hub integration ────────────────────────────────────────────────────────
# hub_dir: ~/.crucible-hub                # cross-project knowledge store
# active_track: ""                        # current research track

# ─── Sync exclusions ────────────────────────────────────────────────────────
# Paths excluded when rsyncing code to fleet pods. Keep this list tight —
# anything not listed here gets shipped with every sync.
sync_excludes:
  - .git
  - .venv
  - __pycache__
  - .crucible/designs
  - .crucible/context
  - .crucible/notes
  - .crucible/store.jsonl
  - .crucible/notes.jsonl
  - logs
  - data/datasets
  - fleet_runs
  - .DS_Store

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