Functions — github.com/OpenPipe/ART

hub / github.com/OpenPipe/ART / functions

Functions1,502 in github.com/OpenPipe/ART

⨍ Functions 1,502◇ Types & classes 261↳ Endpoints 14

↓ 2 callersFunction_tool_call_payload

Minimal JSON payload for embedding in <tool_call> blocks.

src/art/tinker/cookbook_v/renderers/base.py:623

↓ 2 callersMethod_train_model

( self, model: AnyTrainableModel, trajectory_groups: list[TrajectoryGroup], co

src/art/backend.py:45

↓ 2 callersMethod_training_stage

(self)

src/art/pipeline_trainer/trainer.py:374

↓ 2 callersMethod_warn_if_user_system_message

Warn if user provides system message when use_system_prompt=True.

src/art/tinker/cookbook_v/renderers/gpt_oss.py:375

↓ 2 callersFunction_wrap_text_to_width

Greedy word-wrapping that ensures each line fits within `max_width`.

dev/yes-no-maybe-vision/generate_images.py:54

↓ 2 callersMethodactivate_context

(self)

src/art/metrics.py:227

↓ 2 callersFunctionapply_agent_move

(game: TwentyFortyEightGame, move_xml: str)

examples/2048/utils.py:144

↓ 2 callersFunctionauto_migrate_on_register

Automatically migrate any JSONL files found when a model is registered. This is called by the backend during model registration to ensure

src/art/utils/trajectory_migration.py:454

↓ 2 callersFunctionbenchmark_comparison_model

(comparison_model: art.Model)

examples/tic_tac_toe/tic-tac-toe.py:118

↓ 2 callersFunctionbuild_data_metrics_from_summary

( summary: TrajectoryBatchSummary, *, include_trainable_groups: bool, )

src/art/metrics_taxonomy.py:79

↓ 2 callersFunctionbuild_scenario

()

src/art/pipeline_trainer/binary_prefix_tool_pipeline.py:307

↓ 2 callersFunctioncalculate_logprobs

( dtype_for_autocast: torch.dtype, trainer: "GRPOTrainer", input_ids: torch.Tensor, causal_mas

src/art/unsloth/train.py:320

↓ 2 callersFunctioncalculate_rewards

( prompts, completions, rows, )

examples/hn_title_generator/reference_grpo_trainer.py:136

↓ 2 callersFunctioncalculate_step_std_dev

(trajectory_groups: list[TrajectoryGroup])

src/art/utils/old_benchmarking/calculate_step_metrics.py:7

↓ 2 callersFunctionconvert_checkpoint_if_needed

Convert a checkpoint's MoE LoRA adapter to per-expert format if needed. This is a no-op for non-MoE adapters.

src/art/utils/convert_moe_lora.py:144

↓ 2 callersFunctionconvert_litellm_choice_to_openai

( litellm_choice: Choices | StreamingChoices, )

src/art/utils/litellm.py:20

↓ 2 callersFunctioncreate_train_inputs

Create TrainInputs for a single batch offset.

src/art/preprocessing/inputs.py:19

↓ 2 callersMethoddecode

(self, tokens: int | list[int])

tests/unit/test_tinker_renderers.py:47

↓ 2 callersFunctiondo_rollout

Core rollout logic used by both training and eval.

src/art/pipeline_trainer/binary_prefix_tool_pipeline.py:232

↓ 2 callersFunctionensure_list

Normalize content to list form. Wraps string content in a TextPart.

src/art/tinker/cookbook_v/renderers/base.py:417

↓ 2 callersFunctionfilter_rename_model_split

( df: pl.DataFrame, models: list[BenchmarkModelKey] )

src/art/utils/benchmarking/filter_model_split.py:12

↓ 2 callersMethodformat_docstring

(self, indent: str)

src/art/tinker/cookbook_v/renderers/kimi_k2_5_tool_declaration_ts.py:71

↓ 2 callersMethodfrom_dicts

Create a GeneratedScenarioCollection from a list of dictionaries.

src/art/mcp/types.py:104

↓ 2 callersFunctiongenerate_game

(board_length: int = 4)

examples/2048/utils.py:43

↓ 2 callersFunctionget_base_model

Get the base model to use for testing.

tests/integration/test_multi_checkpoint_training.py:33

↓ 2 callersFunctionget_checkpoint_step_not_after_from_s3

Get the latest checkpoint step number that is not after the specified step from S3. Args: not_after_step: Find the latest checkpoint

src/art/utils/s3_checkpoint_utils.py:54

↓ 2 callersMethodget_difficulty_distribution

Get the distribution of difficulties.

src/art/mcp/types.py:169

↓ 2 callersFunctionget_free_port

Returns the first free port >= 8000.

src/art/tinker/server.py:496

↓ 2 callersFunctionget_latest_artifact_provenance

Fetch provenance from the latest W&B artifact's metadata.

tests/integration/test_provenance.py:40

↓ 2 callersFunctionget_llm

Create an AsyncLLM engine with model download and patches applied. Args: args: The engine arguments including model name and configu

src/art/vllm/engine.py:17

↓ 2 callersFunctionget_log_fn

( trainer: Any, results_queue: asyncio.Queue[dict[str, float]] )

src/art/unsloth/train.py:222

↓ 2 callersFunctionget_models_dir

(project_name: str, art_path: str | None = None)

src/art/utils/output_dirs.py:12

↓ 2 callersFunctionget_opponent_move

(game: TicTacToeGame)

examples/tic_tac_toe/game_utils.py:39

↓ 2 callersFunctionget_repo_root_path

()

src/art/utils/get_repo_root_path.py:4

↓ 2 callersFunctionget_tokenizer

Get a tokenizer by name. Checks custom registry first, then falls back to HuggingFace AutoTokenizer.

src/art/tinker/cookbook_v/tokenizer_utils.py:71

↓ 2 callersFunctionget_worker

Get the current worker instance

src/art/vllm/engine.py:72

↓ 2 callersMethodhandle_httpx_response

(self, response: httpx._models.Response)

src/art/auto_trajectory.py:82

↓ 2 callersFunctionimage_to_chunk

Convert a PIL Image to a tinker.types.ImageChunk for QwenVL

src/art/tinker/cookbook_v/renderers/base.py:1224

↓ 2 callersFunctionis_cumulative_metric_key

(key: str)

src/art/metrics.py:27

↓ 2 callersFunctionload_data

Loads, preprocesses, and filters the dataset.

examples/hn_title_generator/train.py:76

↓ 2 callersMethodload_lora

(self, adapter_model: dict[str, torch.Tensor])

src/art/megatron/lora.py:63

↓ 2 callersFunctionload_title_data

Loads and prepares data specifically for the title generation task.

examples/hn_title_generator/reference_grpo_trainer.py:81

↓ 2 callersFunctionload_trajectories

Load and flatten trajectory files (Parquet) into a Polars DataFrame. The expected on-disk layout is:: {art_path}/{project_name}/mod

src/art/utils/benchmarking/load_trajectories.py:27

↓ 2 callersMethodload_weights

( self, adapter_model: dict[str, torch.Tensor], *, suffix: str, into:

src/art/megatron/lora.py:79

↓ 2 callersMethodlog_if_due

(self, *, force: bool = False)

src/art/pipeline_trainer/status.py:93

↓ 2 callersFunctionmax_cell_value

(game: TwentyFortyEightGame)

examples/2048/utils.py:162

↓ 2 callersMethodmeasure

(self, key: str)

src/art/metrics.py:186

↓ 2 callersFunctionmessage_or_choice_to_dict

(message_or_choice: MessageOrChoice)

src/art/utils/trajectory_migration.py:79

↓ 2 callersMethodnote_group_dequeued

(self, _group: TrajectoryGroup)

src/art/pipeline_trainer/status.py:117

↓ 2 callersMethodnote_training_end

(self)

src/art/pipeline_trainer/status.py:136

↓ 2 callersFunctionoffload_to_cpu

Offload model params and optimizer state to CPU pinned memory.

src/art/megatron/offload.py:14

↓ 2 callersFunctionopenai_server_task

Starts an asyncio task that runs an OpenAI-compatible server. Args: engine: The vLLM engine client. config: The configuratio

src/art/vllm/server.py:22

↓ 2 callersMethodoverwrite_state

Overwrite persistent state in the model directory as JSON. This state is stored in `state.json` within the model's output directory a

src/art/model.py:311

↓ 2 callersFunctionpatch_listen_for_disconnect

()

src/art/vllm/patches.py:22

↓ 2 callersFunctionprocess_train_batch

Process training batches and yield results. Yields tuples of (result, warmup_done) where warmup_done indicates if warmup just finished.

src/art/unsloth/service.py:90

↓ 2 callersFunctionprompt_for_title

(content: str)

examples/hn_title_generator/utils.py:146

↓ 2 callersFunctionpull_data

( split: str = "train", max_items: int = 10, min_score: int = 20, )

examples/hn_title_generator/utils.py:131

↓ 2 callersMethodreload_to_gpu

Reload training model and optimizer back to GPU using async transfers.

src/art/unsloth/service.py:270

↓ 2 callersMethodrender_message

Render a single message to tokens. Args: message: The message to render. ctx: Context about the message's position, i

src/art/tinker/cookbook_v/renderers/deepseek_v3.py:66

↓ 2 callersMethodrender_message

Render a message. For assistant messages, ctx.is_last controls whether thinking is preserved (True) or stripped to empty <think></thi

src/art/tinker/cookbook_v/renderers/kimi_k2.py:406

↓ 2 callersFunctionreward_for_answer

(content: str | None)

dev/yes-no-maybe-metrics.py:61

↓ 2 callersFunctionreward_for_answer

Score: maybe=1.0, no=0.75, yes=0.5, other=0.0.

src/art/pipeline_trainer/yes_no_maybe_pipeline.py:46

↓ 2 callersFunctionrl_rollout

Single RL rollout with reward based on response.

dev/sft/sft-warmup.py:24

↓ 2 callersFunctionrollout

( client: openai.AsyncOpenAI, model: art.TrainableModel, prompt: str, *, max_tokens: int,

dev/yes-no-maybe-metrics.py:94

↓ 2 callersFunctionrollout

(client: openai.AsyncOpenAI, prompt: str)

dev/yes-no-maybe.py:12

↓ 2 callersFunctionrollout

(scenario: Scenario)

dev/math-vista/math-vista.py:42

↓ 2 callersFunctionrollout

(model: art.Model[MyConfig], scenario: MyTask)

examples/benchmarking_comparison_models.py:84

↓ 2 callersFunctionrollout

(model: art.Model, puzzle: TemporalCluePuzzle)

examples/temporal_clue/temporal-clue.py:35

↓ 2 callersFunctionrollout

Generates a title, validates it, scores it, and returns a trajectory.

examples/hn_title_generator/train.py:159

↓ 2 callersFunctionsay_hi

A method that says hi to an assistant and returns the response.

tests/unit/test_auto_trajectory.py:229

↓ 2 callersFunctionscenario_id_for_prompt

(prompt: str)

dev/yes-no-maybe-metrics.py:71

↓ 2 callersFunctionscrape_article

Scrape article text from a news URL (Fox News, MSNBC, etc.) Args: url: The URL of the news article to scrape Returns: T

examples/just-the-facts/just_the_facts/utils.py:9

↓ 2 callersFunctionsimple_rollout

Minimal rollout that produces a single turn with a reward.

tests/integration/test_provenance.py:15

↓ 2 callersFunctionstreamlined_traceback

()

src/mp_actors/traceback.py:33

↓ 2 callersFunctiontotal_actor_tokens

(groups: list[art.TrajectoryGroup])

dev/yes-no-maybe-metrics.py:86

↓ 2 callersFunctiontrain

( trainer: "GRPOTrainer", results_queue: asyncio.Queue[dict[str, float]], )

src/art/unsloth/train.py:60

↓ 2 callersMethodtrain

( self, disk_packed_tensors: DiskPackedTensors, config: types.TrainConfig, _co

src/art/unsloth/service.py:560

↓ 1 callersMethod__init__

( self, trajectories: ( Iterable[Trajectory | BaseException] | Iterable[Awaitable[

src/art/trajectories.py:137

↓ 1 callersMethod_add_default_step_metrics

( self, trajectory_groups: list[TrajectoryGroup], *, split: str, provi

src/art/model.py:637

↓ 1 callersMethod_allocated_gpu_count

(self, model: Model)

src/art/local/backend.py:147

↓ 1 callersFunction_anthropic_message

(*, api_key: str, prefix: str)

tests/integration/test_live_api_cost.py:64

↓ 1 callersMethod_apply_policy_versions

( self, group: TrajectoryGroup, *, initial_version: int, final_version

src/art/pipeline_trainer/trainer.py:653

↓ 1 callersMethod_apply_scenario_metadata

( self, group: TrajectoryGroup, scenario: ScenarioT )

src/art/pipeline_trainer/trainer.py:666

↓ 1 callersMethod_assistant_header_suffix

Additional assistant header text injected before content.

src/art/tinker/cookbook_v/renderers/qwen3.py:440

↓ 1 callersMethod_average_steps_off_policy

( self, current_step: int, batch: list[TrajectoryGroup] )

src/art/pipeline_trainer/trainer.py:748

↓ 1 callersMethod_build_model_state

(self, model: TrainableModel)

src/art/tinker_native/backend.py:527

↓ 1 callersMethod_build_models_dict

Build models dict from checkpoint info files.

src/art/tinker/service.py:233

↓ 1 callersFunction_build_parser

()

scripts/ci/compute_uv_fingerprint.py:17

↓ 1 callersFunction_calculate_logprobs

( lm_head_t: torch.Tensor, # Shape [H, V] hidden_states: torch.Tensor, # Shape [B, S, H] next_in

src/art/unsloth/train.py:356

↓ 1 callersFunction_canonicalize_upstream_metric_key

(metric: str)

src/art/serverless/backend.py:60

↓ 1 callersFunction_canonicalize_upstream_metrics

(metrics: dict[str, float])

src/art/unsloth/train.py:52

↓ 1 callersMethod_collect_batch

( self, current_step: int )

src/art/pipeline_trainer/trainer.py:493

↓ 1 callersMethod_compute_batch_avg_reward

(self, batch: list[TrajectoryGroup])

src/art/pipeline_trainer/status.py:353

↓ 1 callersMethod_compute_batch_avg_std_dev

(self, batch: list[TrajectoryGroup])

src/art/pipeline_trainer/status.py:347

↓ 1 callersMethod_compute_group_std_dev

(self, group: TrajectoryGroup)

src/art/pipeline_trainer/status.py:339

↓ 1 callersMethod_compute_rollups

(self, cost_metrics: dict[str, float])

src/art/metrics.py:347

↓ 1 callersFunction_configured_token_pricing

(model_name: str)

src/art/api_costs.py:60

↓ 1 callersMethod_consume_batch_rollout_timings

( self, batch: list[TrajectoryGroup] )

src/art/pipeline_trainer/trainer.py:797

↓ 1 callersFunction_convert_job_status

(status: str, message: str | None = None)

src/art/utils/deployment/together.py:114

↓ 1 callersMethod_count_zero_variance_groups

(self, batch: list[TrajectoryGroup])

src/art/pipeline_trainer/status.py:329

← previous next →301–400 of 1,502, ranked by callers