`dockyard_rl.data.processors`¶

Contains data processors for training and evaluation.

Module Contents¶

Functions¶

`helpsteer3_data_processor`	Process a HelpSteer3 preference datum into a DatumSpec for GRPO training.
`sft_processor`	Process a datum dictionary for SFT training.
`preference_preprocessor`	Process a datum dictionary for RM/DPO training.
`text_preprocessor`	Tokenize a single text prompt into a DatumSpec for evaluation rollouts.
`math_data_processor`	Process a datum dictionary (directly loaded from dataset) into a DatumSpec for the Math Environment.
`math_hf_data_processor`	Process a datum dictionary (directly loaded from data/hf_datasets/openmathinstruct2.py) into a DatumSpec for the Reward Model Environment.
`vlm_hf_data_processor`	Process a datum dictionary (directly loaded from response_datasets/<dataset_name>.py) into a DatumSpec for the VLM Environment.
`multichoice_qa_processor`	Process a datum dictionary (directly loaded from dataset) into a DatumSpec for multiple-choice problems.
`swe_bench_data_processor`	Process a SWE-bench row (from `SWEBenchDataset.format_data`) into a DatumSpec.
`swe_bench_pro_data_processor`	Process a SWE-bench Pro row (from `SWEBenchProDataset.format_data`).
`terminal_bench_data_processor`	Process a Terminal-Bench row (from `TerminalBenchDataset._build_row`).
`program_bench_data_processor`	Process a ProgramBench row (from `ProgramBenchDataset._build_row`).
`hle_data_processor`	Process an HLE row (from `HLEDataset.format_data`) into a DatumSpec.
`gdpval_data_processor`	Process a GDPval row (from `GDPvalDataset.format_data`) into a DatumSpec.
`gdpval_agentic_data_processor`	Process a GDPval agentic row (from `GDPvalAgenticDataset.format_data`).
`gym_data_processor`	Process a datum dictionary (directly loaded from dataset) into a DatumSpec for Gym.
`register_processor`

Data¶

`TokenizerType`
`PROCESSOR_REGISTRY`

API¶

dockyard_rl.data.processors.TokenizerType¶: None

dockyard_rl.data.processors.helpsteer3_data_processor(datum_dict: dict[str, Any], task_data_spec: dockyard_rl.data.interfaces.TaskDataSpec, tokenizer: dockyard_rl.data.processors.TokenizerType, max_seq_length: int, idx: int) → dockyard_rl.data.interfaces.DatumSpec¶

Process a HelpSteer3 preference datum into a DatumSpec for GRPO training.

This function converts HelpSteer3 preference data to work with GRPO by:

Using the context as the prompt
Using the preferred completion as the target response
Creating a reward signal based on preference scores

dockyard_rl.data.processors.sft_processor(datum_dict: dict[str, Any], task_data_spec: dockyard_rl.data.interfaces.TaskDataSpec, tokenizer, max_seq_length: int, idx: int, add_bos: bool = True, add_eos: bool = True, add_generation_prompt: bool = False) → dockyard_rl.data.interfaces.DatumSpec¶: Process a datum dictionary for SFT training.

dockyard_rl.data.processors.preference_preprocessor(datum_dict: dict[str, Any], task_data_spec: dockyard_rl.data.interfaces.TaskDataSpec, tokenizer, max_seq_length: int, idx: int) → dockyard_rl.data.interfaces.PreferenceDatumSpec¶: Process a datum dictionary for RM/DPO training.

dockyard_rl.data.processors.text_preprocessor(datum_dict: dict[str, Any], task_data_spec: dockyard_rl.data.interfaces.TaskDataSpec, tokenizer: dockyard_rl.data.processors.TokenizerType, max_input_seq_length: int, idx: int = 0) → dockyard_rl.data.interfaces.DatumSpec¶

Tokenize a single text prompt into a DatumSpec for evaluation rollouts.

The prompt text is datum_dict["input"]; the optional datum_dict["ground_truth"] is threaded into extra_env_info so the scoring environment can verify the generated answer. Mirrors the prompt construction of math_data_processor so eval samples collate and roll out identically to training samples.

dockyard_rl.data.processors.math_data_processor(datum_dict: dict[str, Any], task_data_spec: dockyard_rl.data.interfaces.TaskDataSpec, tokenizer: dockyard_rl.data.processors.TokenizerType, max_seq_length: int, idx: int) → dockyard_rl.data.interfaces.DatumSpec¶: Process a datum dictionary (directly loaded from dataset) into a DatumSpec for the Math Environment.

dockyard_rl.data.processors.math_hf_data_processor(datum_dict: dict[str, Any], task_data_spec: dockyard_rl.data.interfaces.TaskDataSpec, tokenizer: dockyard_rl.data.processors.TokenizerType, max_seq_length: int, idx: int) → dockyard_rl.data.interfaces.DatumSpec¶: Process a datum dictionary (directly loaded from data/hf_datasets/openmathinstruct2.py) into a DatumSpec for the Reward Model Environment.

dockyard_rl.data.processors.vlm_hf_data_processor(datum_dict: dict[str, Any], task_data_spec: dockyard_rl.data.interfaces.TaskDataSpec, processor: transformers.AutoProcessor, max_seq_length: int, idx: int) → dockyard_rl.data.interfaces.DatumSpec¶: Process a datum dictionary (directly loaded from response_datasets/<dataset_name>.py) into a DatumSpec for the VLM Environment.

dockyard_rl.data.processors.multichoice_qa_processor(datum_dict: dict[str, Any], task_data_spec: dockyard_rl.data.interfaces.TaskDataSpec, tokenizer: dockyard_rl.data.processors.TokenizerType, max_seq_length: int, idx: int) → dockyard_rl.data.interfaces.DatumSpec¶: Process a datum dictionary (directly loaded from dataset) into a DatumSpec for multiple-choice problems.

dockyard_rl.data.processors.swe_bench_data_processor(datum_dict: dict[str, Any], task_data_spec: dockyard_rl.data.interfaces.TaskDataSpec, tokenizer: dockyard_rl.data.processors.TokenizerType, max_seq_length: int | None, idx: int) → dockyard_rl.data.interfaces.DatumSpec¶

Process a SWE-bench row (from SWEBenchDataset.format_data) into a DatumSpec.

The prompt is the system + user message log with no answer turn (the agent generates the fix at rollout time). The held-out test metadata is routed into extra_env_info so the reward function can verify the solution.

dockyard_rl.data.processors.swe_bench_pro_data_processor(datum_dict: dict[str, Any], task_data_spec: dockyard_rl.data.interfaces.TaskDataSpec, tokenizer: dockyard_rl.data.processors.TokenizerType, max_seq_length: int | None, idx: int) → dockyard_rl.data.interfaces.DatumSpec¶

Process a SWE-bench Pro row (from SWEBenchProDataset.format_data).

The prompt is the system + user message log with no answer turn. The held-out scoring inputs — the per-instance image, harness scripts, the gold test checkout command, the selected test files, the named fail/pass tests, and the gold test/fix diffs — are routed into extra_env_info so the image-mode reward can score the solution without exposing them to the agent.

dockyard_rl.data.processors.terminal_bench_data_processor(datum_dict: dict[str, Any], task_data_spec: dockyard_rl.data.interfaces.TaskDataSpec, tokenizer: dockyard_rl.data.processors.TokenizerType, max_seq_length: int | None, idx: int) → dockyard_rl.data.interfaces.DatumSpec¶

Process a Terminal-Bench row (from TerminalBenchDataset._build_row).

The prompt is the system + user message log with no answer turn. Everything the multi-turn environment needs — the image/Dockerfile + build context, the held-out tests/ (injected only at finish), the verifier command/mount/ result paths, timeouts, container env, network policy, and turn budget — is routed into extra_env_info. The dict-valued fields are carried as JSON strings in the dataset row (Arrow cannot hold variable-key structs) and decoded back here.

dockyard_rl.data.processors.program_bench_data_processor(datum_dict: dict[str, Any], task_data_spec: dockyard_rl.data.interfaces.TaskDataSpec, tokenizer: dockyard_rl.data.processors.TokenizerType, max_seq_length: int | None, idx: int) → dockyard_rl.data.interfaces.DatumSpec¶

Process a ProgramBench row (from ProgramBenchDataset._build_row).

The prompt is the system + user message log with no answer turn. The image ref, the held-out test branches (base64 tars + expected node IDs, carried as a JSON string since Arrow cannot hold variable-key structs), language, and the per-episode budgets are routed into extra_env_info for the multi-turn environment and the clean-rebuild grading reward.

dockyard_rl.data.processors.hle_data_processor(datum_dict: dict[str, Any], task_data_spec: dockyard_rl.data.interfaces.TaskDataSpec, tokenizer: dockyard_rl.data.processors.TokenizerType, max_seq_length: int | None, idx: int) → dockyard_rl.data.interfaces.DatumSpec¶

Process an HLE row (from HLEDataset.format_data) into a DatumSpec.

The prompt is the system (HLE response format) + user (question) message log with no answer turn. The gold answer, the question (needed by the LLM judge), and the answer type/category are routed into extra_env_info for the HLE verifier environment.

dockyard_rl.data.processors.gdpval_data_processor(datum_dict: dict[str, Any], task_data_spec: dockyard_rl.data.interfaces.TaskDataSpec, tokenizer: dockyard_rl.data.processors.TokenizerType, max_seq_length: int | None, idx: int) → dockyard_rl.data.interfaces.DatumSpec¶

Process a GDPval row (from GDPvalDataset.format_data) into a DatumSpec.

The prompt is the system (text-deliverable instructions) + user (task prompt) message log with no answer turn. The task prompt and rubric_json (needed by the rubric judge) plus occupation/sector are routed into extra_env_info.

dockyard_rl.data.processors.gdpval_agentic_data_processor(datum_dict: dict[str, Any], task_data_spec: dockyard_rl.data.interfaces.TaskDataSpec, tokenizer: dockyard_rl.data.processors.TokenizerType, max_seq_length: int | None, idx: int) → dockyard_rl.data.interfaces.DatumSpec¶

Process a GDPval agentic row (from GDPvalAgenticDataset.format_data).

The prompt is the system (tool-use instructions) + user (task prompt) message log with no answer turn. Everything the multi-turn file-producing environment needs — the container image, the deliverable/reference dirs, the turn/exec budgets, the task prompt and rubric_json for the judge, and any base64 reference files (carried as a JSON string since Arrow cannot hold variable-key structs) — is routed into extra_env_info.

dockyard_rl.data.processors.gym_data_processor(datum_dict: dict[str, Any], task_data_spec: dockyard_rl.data.interfaces.TaskDataSpec, tokenizer: dockyard_rl.data.processors.TokenizerType, max_seq_length: int | None, idx: int) → dockyard_rl.data.interfaces.DatumSpec¶

Process a datum dictionary (directly loaded from dataset) into a DatumSpec for Gym.

Note: this processor is kept for registry compatibility but Gym integration is not a Dockyard target. The processor is deferred-importable without side effects.

dockyard_rl.data.processors.PROCESSOR_REGISTRY: Dict[str, dockyard_rl.data.interfaces.TaskDataProcessFnCallable]¶: ‘cast(…)’

dockyard_rl.data.processors.register_processor(processor_name: str, processor_function: dockyard_rl.data.interfaces.TaskDataProcessFnCallable) → None¶

dockyard_rl.data.processors¶

Module Contents¶

Functions¶

Data¶

API¶

`dockyard_rl.data.processors`¶