README.md

---
dataset_info:
  features:
  - name: data_source
    dtype: string
  - name: prompt
    list:
    - name: content
      dtype: string
    - name: role
      dtype: string
  - name: ability
    dtype: string
  - name: reward_model
    struct:
    - name: ground_truth
      dtype: 'null'
    - name: style
      dtype: string
  - name: extra_info
    struct:
    - name: split
      dtype: string
    - name: tools_kwargs
      struct:
      - name: difficulty
        dtype: int64
      - name: idx
        dtype: int64
      - name: item
        struct:
        - name: add_info
          struct:
          - name: had_tests
            dtype: bool
        - name: class_code
          dtype: string
        - name: class_code_updated
          dtype: string
        - name: class_doc
          dtype: string
        - name: class_id
          dtype: int64
        - name: class_loc
          dtype: int64
        - name: class_name
          dtype: string
        - name: class_parent
          dtype: string
        - name: class_position_in_file
          dtype: int64
        - name: class_without_method
          dtype: string
        - name: difficulty
          dtype: int64
        - name: dp_id
          dtype: string
        - name: file_imports
          dtype: string
        - name: file_level_code
          dtype: string
        - name: file_path
          dtype: string
        - name: file_without_method
          dtype: string
        - name: idx
          dtype: int64
        - name: line_numbers
          dtype: string
        - name: method
          struct:
          - name: body
            dtype: string
          - name: declaration
            dtype: string
          - name: description
            dtype: string
          - name: global_method_body_index
            list: int64
          - name: global_method_declaration_index
            list: int64
          - name: method_body_index
            list: int64
          - name: method_declaration_index
            list: int64
          - name: name
            dtype: string
          - name: num_lines
            dtype: int64
          - name: num_lines_quartile
            dtype: int64
        - name: method_count
          dtype: int64
        - name: qwen3-8b_pass_repair
          dtype: bool
        - name: qwen3-8b_pass_single
          dtype: bool
        - name: raw_doc
          dtype: string
        - name: raw_file_content
          dtype: string
        - name: split
          dtype: string
        - name: tests
          struct:
          - name: full_paths
            list: string
          - name: methods
            list: string
          - name: modules
            list: string
          - name: test_output
            dtype: string
          - name: time_cat
            dtype: float64
          - name: time_edit
            dtype: float64
          - name: time_reset
            dtype: float64
          - name: time_reset_norm
            dtype: float64
          - name: time_test
            dtype: float64
          - name: time_test_norm
            dtype: float64
          - name: time_total
            dtype: float64
  - name: agent_name
    dtype: string
  splits:
  - name: train
    num_bytes: 232713631
    num_examples: 1364
  - name: test
    num_bytes: 14568916
    num_examples: 100
  download_size: 247410221
  dataset_size: 247282547
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
---

`JetBrains-Research/django_method_gen` is a code generation benchmark built from the Django codebase.

This dataset is used in the example of [IDEGym](https://github.com/JetBrains-Research/idegym/) usage for VERL-based RL training (see [`this repo`](https://github.com/JetBrains-Research/idegym/tree/main/examples/verl)).

Each example is a task to regenerate a single Python method that has been cut from its class. The dataset provides the surrounding class code, file imports, and docstrings as context. Reward is rule-based: the agent's submission is evaluated by running the original unit tests.

The dataset follows the VERL multi-turn format: each row contains a `prompt` (chat-style system + user messages), an `agent_name` field (`"idegym_django"`), and an `extra_info` blob carrying the raw task data passed to the IDEGym server — including the method body to recover, file context, and test metadata. There are 1,364 training examples and 100 test examples, spanning four difficulty levels.