CoolFace
Datasetpublic

ronaldcmz/Claude-Opus-Dataclaw-Unredacted

Claude Opus Dataclaw Unredacted How this dataset was built Collected the local Petromallet raw export plus selected public Dataclaw uploads. Filtered to the supported Opus-family source rows. Deduplicated by session_id and first user message. Converted raw assistant tool_uses directly into structured OpenAI-style tool_calls. Derived per-row tool definitions from canonical schemas and observed tool usage. Preserved assistant reasoning in <think>...</think> blocks.… See the full description on the dataset page: https://huggingface.co/datasets/ronaldcmz/Claude-Opus-Dataclaw-Unredacted.

sourceHugging Facemitupdated 3mo agoView on Hugging Face
0likes27downloads
Dataset Card

Claude Opus Dataclaw Unredacted

How this dataset was built

  1. 1.Collected the local Petromallet raw export plus selected public Dataclaw uploads.
  2. 2.Filtered to the supported Opus-family source rows.
  3. 3.Deduplicated by session_id and first user message.
  4. 4.Converted raw assistant tool_uses directly into structured OpenAI-style tool_calls.
  5. 5.Derived per-row tool definitions from canonical schemas and observed tool usage.
  6. 6.Preserved assistant reasoning in <think>...</think> blocks.
  7. 7.Removed Claude Code wrapper messages such as <local-command-stdout>...</local-command-stdout>.
  8. 8.Ran an AI redaction-filling pass over retained rows.
  9. 9.Validated the final JSONL structure.

This release does not synthesize missing tool-result messages. The source data mostly contains user/assistant turns plus assistant tool calls, so the rebuilt output keeps that structure.

Output Schema

Each JSONL row contains the following top-level fields:

FieldDescription
promptThe first user message, included for convenience
messagesConversation turns in OpenAI chat format
toolsTool definitions in OpenAI function-calling format
original_modelThe source model label from the original dataset row
source_datasetThe local raw source identifier or Hugging Face dataset repo ID
metadataOriginal metadata such as session_id, project, branch, timing, and stats

Message Format

Assistant messages with tool calls use native structured arguments:

json
{
  "role": "assistant",
  "content": "<think>Need to inspect the file first.</think>",
  "tool_calls": [
    {
      "type": "function",
      "id": "call_abc123",
      "function": {
        "name": "run_command",
        "arguments": {
          "command": "ls -la"
        }
      }
    }
  ]
}

This release contains no `tool` role messages because the direct rebuild does not invent missing tool outputs.

Tool Definition Format

json
{
  "type": "function",
  "function": {
    "name": "run_command",
    "description": "Run a bash command.",
    "parameters": {
      "type": "object",
      "properties": {
        "command": {"type": "string"}
      },
      "required": ["command"]
    }
  }
}

Final Stats

  • 436 final rows
  • 58,077 total messages
  • 5,785 user messages
  • 52,292 assistant messages
  • 428 rows with tool definitions
  • 428 rows with assistant tool calls
  • 48,779 total tool calls
  • 0 remaining literal `[REDACTED]` occurrences

Source Breakdown