Skip to content

CONTRIBUTING.md

Latest commit

 

History

History
96 lines (77 loc) · 2.93 KB

CONTRIBUTING.md

File metadata and controls

96 lines (77 loc) · 2.93 KB

Contributing to SeevoMap

How to Contribute a Node

Every node in SeevoMap represents a real auto-research execution record. Contributing requires just 3 fields:

Minimum Node (3 JSON fields)

{
  "task": {"domain": "pretraining"},
  "idea": {"text": "Replace RMSNorm with LayerNorm for faster convergence"},
  "result": {"metric_name": "val_loss", "metric_value": 3.18}
}

Submit via CLI:

seevomap submit my_experiment.json

Or via SDK:

from seevomap import SeevoMap
svm = SeevoMap()
svm.submit({"task": {"domain": "pretraining"}, ...})

Submitted nodes go through a review process before appearing in the public graph.

Full Node (Recommended)

Providing more detail makes your experience more useful to others:

{
  "task": {
    "domain": "pretraining",
    "name": "nanogpt-speedrun",
    "description": "Minimize val_loss for GPT-2 124M on FineWeb within 1 hour"
  },
  "idea": {
    "text": "[Experiment] Replace RMSNorm with LayerNorm...\n[Code Changes] In model.py, change...\n[End]",
    "method_tags": ["normalization", "architecture"]
  },
  "result": {
    "metric_name": "val_loss",
    "metric_value": 3.18,
    "baseline_value": 3.255,
    "success": true
  },
  "context": {
    "model": "claude-opus-4-6",
    "epoch": 3,
    "source": "my-autoresearch-project",
    "hardware": "4xA100",
    "date": "2026-03-21T10:00:00Z"
  },
  "code_diff": "--- a/model.py\n+++ b/model.py\n@@ ...",
  "config": {"num_layers": 12, "lr": 0.001},
  "analysis": "LayerNorm converges faster in first 500 steps but similar final loss. Main benefit is training stability."
}

Field Reference

Field Required Description
task.domain Yes Research domain (e.g., pretraining, posttraining, chemistry)
idea.text Yes What was tried. Recommend [Experiment]...[Code Changes]...[End] format
result.metric_name + result.metric_value Yes Execution result
idea.method_tags Recommended Method tags for structured search
result.baseline_value Recommended Baseline comparison
result.success Recommended Whether it beat the baseline
code_diff Optional Unified diff — greatly improves reproducibility
config Optional Hyperparameter dict
analysis Optional Why it succeeded or failed
context Optional Execution environment info

How Your Node Gets Used

  1. Your node joins the graph and gets an embedding computed automatically
  2. When another agent searches for related experiences, your node may be matched
  3. The matched node's idea.text + result are injected into the agent's prompt
  4. The agent learns from your experience (successes to build on, failures to avoid)
  5. New execution results become new nodes — the network gets smarter with every use

Batch Submit

Submit a directory of experiment JSON files:

seevomap submit --dir ./my_experiments/