unsloth/studio/frontend/src/features/data-recipes/learning-recipes/text-to-python.json

{
  "recipe": {
    "model_providers": [
      {
        "name": "provider_1",
        "endpoint": "https://openrouter.ai/api/v1",
        "provider_type": "openai",
        "extra_headers": {},
        "extra_body": {}
      }
    ],
    "mcp_providers": [],
    "model_configs": [
      {
        "alias": "model_1",
        "model": "mistralai/ministral-8b-2512",
        "provider": "provider_1",
        "inference_parameters": {
          "temperature": 0.7,
          "max_tokens": 2048
        }
      }
    ],
    "tool_configs": [],
    "columns": [
      {
        "column_type": "sampler",
        "name": "domain",
        "drop": false,
        "sampler_type": "category",
        "params": {
          "values": [
            "Data Processing",
            "Web API",
            "Automation"
          ]
        }
      },
      {
        "column_type": "sampler",
        "name": "task_type",
        "drop": false,
        "sampler_type": "subcategory",
        "params": {
          "category": "domain",
          "values": {
            "Data Processing": [
              "CSV cleaning",
              "JSON transform",
              "deduplicate rows"
            ],
            "Web API": [
              "GET endpoint",
              "POST validation",
              "pagination helper"
            ],
            "Automation": [
              "file organizer",
              "log parser",
              "daily report script"
            ]
          }
        }
      },
      {
        "column_type": "llm-text",
        "name": "instruction",
        "drop": false,
        "model_alias": "model_1",
        "prompt": "Write one clear Python coding instruction.\nDomain: {{ domain }}\nTask type: {{ task_type }}\n\nKeep it practical and specific.\nReturn only the instruction without any code.",
        "with_trace": "none"
      },
      {
        "column_type": "llm-code",
        "name": "code_implementation",
        "drop": false,
        "model_alias": "model_1",
        "prompt": "Write Python code for:\n{{ instruction }}\n\nRequirements:\n- runnable script or function\n- include needed imports\n- short comments only where useful\n- no markdown fences",
        "code_lang": "python"
      },
      {
        "column_type": "llm-judge",
        "name": "code_judge_result",
        "drop": false,
        "model_alias": "model_1",
        "prompt": "Evaluate generated Python code against the instruction.\n\nInstruction:\n{{ instruction }}\n\nCode:\n{{ code_implementation }}",
        "scores": [
          {
            "name": "Correctness",
            "description": "Follows instruction and is executable",
            "options": {
              "0": "bad",
              "1": "partial",
              "2": "good",
              "3": "excellent"
            }
          }
        ]
      }
    ],
    "processors": []
  },
  "run": {
    "rows": 5,
    "preview": true,
    "output_formats": [
      "jsonl"
    ]
  },
  "ui": {
    "nodes": [
      {
        "id": "provider_1",
        "x": 1032.6798211423347,
        "y": -450.4885376732656,
        "width": 400
      },
      {
        "id": "model_1",
        "x": 1538.0273166472973,
        "y": -483.2003290046642,
        "width": 400
      },
      {
        "id": "domain",
        "x": 0,
        "y": 24,
        "width": 400
      },
      {
        "id": "task_type",
        "x": 480,
        "y": 24,
        "width": 400
      },
      {
        "id": "instruction",
        "x": 958.8989453654599,
        "y": -9.971266983459952,
        "width": 400
      },
      {
        "id": "code_implementation",
        "x": 1538.788058529745,
        "y": -45.56493974435071,
        "width": 400
      },
      {
        "id": "code_judge_result",
        "x": 2040.9251520522098,
        "y": -13.362336454344792,
        "width": 400
      },
      {
        "id": "note_1",
        "x": 1482.1328175027095,
        "y": 242.4370179053253,
        "width": 568,
        "node_type": "markdown_note",
        "name": "note_1",
        "markdown": "The **LLM Code** block is where Python code is generated from your instruction/prompt.\n\n##### How it works in this recipe:\n\n- You provide a clear prompt (often using Jinja references from earlier columns)\n- The model returns a response\n- The block extracts code content directly for the output column\n\n##### Current status:\n\n- We are **not** running Python lint/syntax validation in this recipe yet (Soon)\n- Validation support is planned and will be added\n\n##### What this means:\n\n- You may get mostly correct code, but some rows can still have syntax/style issues\n- Keep prompts specific and constrained to reduce bad outputs\n\n##### Tip:\n\n- Ask for one self-contained function/script\n- Ask for required imports\n- Ask for no markdown fences if you want cleaner extraction\n",
        "note_color": "#FDE68A",
        "note_opacity": "35"
      },
      {
        "id": "note_2",
        "x": 2513.2527820497985,
        "y": -235.2544980991115,
        "width": 471,
        "node_type": "markdown_note",
        "name": "note_2",
        "markdown": "The **LLM Judge** block evaluates generated outputs with rubric-style scores.\n\n##### Important:\n\n- A judge can have **one or many scores**\n- Each score has:\n  - a name (for example: `Correctness`)\n  - a description\n  - options (value + meaning)\n\n##### Example multi-score setup:\n\n- Correctness\n- Readability\n- Efficiency\n\n##### Why use multiple scores:\n\n- You get richer quality signals than a single pass/fail\n- Easier filtering and weighting later in training data prep\n\n##### Practical pattern:\n\n1. Generate code with LLM Code\n2. Judge with 2-4 focused scores\n3. Keep high-quality rows based on score thresholds\n",
        "note_color": "#FDE68A",
        "note_opacity": "35"
      }
    ],
    "edges": [
      {
        "from": "domain",
        "to": "task_type",
        "type": "canvas",
        "source_handle": "data-out",
        "target_handle": "data-in"
      },
      {
        "from": "task_type",
        "to": "instruction",
        "type": "canvas",
        "source_handle": "data-out",
        "target_handle": "data-in"
      },
      {
        "from": "provider_1",
        "to": "model_1",
        "type": "semantic",
        "source_handle": "semantic-out",
        "target_handle": "semantic-in"
      },
      {
        "from": "instruction",
        "to": "code_implementation",
        "type": "canvas",
        "source_handle": "data-out",
        "target_handle": "data-in"
      },
      {
        "from": "model_1",
        "to": "instruction",
        "type": "semantic",
        "source_handle": "semantic-out-bottom",
        "target_handle": "data-in-top"
      },
      {
        "from": "model_1",
        "to": "code_implementation",
        "type": "semantic",
        "source_handle": "semantic-out-bottom",
        "target_handle": "data-in-top"
      },
      {
        "from": "code_implementation",
        "to": "code_judge_result",
        "type": "canvas",
        "source_handle": "data-out",
        "target_handle": "data-in"
      },
      {
        "from": "model_1",
        "to": "code_judge_result",
        "type": "semantic",
        "source_handle": "semantic-out-bottom",
        "target_handle": "data-in-top"
      }
    ],
    "layout_direction": "LR"
  }
}