A model is an object that can be passed to your `rollout` function, and used to log completions. Additionally, a `TrainableModel`, which is a

Build and accumulate step-level metrics for logging.

Per-million-token pricing for a model.

Model configuration. Args: init: Arguments for initializing an Unsloth FastLanguageModel. engine: Arguments for the vLLM eng

Container for a single turn in a multi-turn conversation. Args: role: Role String that denotes the source of the message, typic

A chunk of text content in a message, usually meant to be visible to the user (unlike ThinkingPart, which is internal reasoning).

Container for parts of a rendered message, structured for loss masking. A rendered message is split into header and output to control which

Represents a tool call that failed to parse from model output. When a model generates text that looks like a tool call but cannot be par

Results from a migration operation.

Represents an MCP tool.

Async 3-stage pipeline for rollouts, training, and eval.

Context passed to render_message for rendering a single message. This allows renderers to access information about the message's position

Structured tool invocation following OpenAI/kosong format. This represents a request to invoke a tool/function. The structure follows th

A collection of scenarios with utilities for management and analysis.

Renderer for OpenAI's open source models using the Harmony format. Wire format: <|start|>role<|channel|>channel<|message|>content<|end|>

A scenario for MCP agent evaluation.

LRU-bounded radix trie for token sequence rewrites.

Represents an MCP resource.

Incremental text content during streaming. Contains only the new text since the last delta, not the accumulated text. The recipient should co

Incremental thinking/reasoning content during streaming. Contains only the new thinking text since the last delta.

Model's internal reasoning (chain-of-thought) as a content part.

Configuration for deploying to Together. See supported base models: https://docs.together.ai/docs/lora-inference#supported-base-models Attri

Renderer for DeepSeek V3 models in NON-THINKING mode. Format: <|begin_of_sentence|><|User|>question<|Assistant|></think>answer<|end_

Result of a deployment operation.

Server configuration. Args: log_file: Path to the log file. server_args: Arguments for the vLLM OpenAI-compatible server.

Emitted at the start of a new message during streaming. This signals that a new message is beginning and provides the author info.

Training inputs with config attached.

Base result returned from backend.train(). Attributes: step: The training step after this training call completed. metrics: Aggre

An error raised when the user attempts to deploy a model to a provider that does not support it for serverless LoRA deployment. Status code:

Client for interacting with Alpha Vantage API

Client for interacting with Balldontlie NBA API

Container for dataset batch information.

Renderer for DeepSeek V3 models in THINKING mode. Format: <|begin_of_sentence|><|User|>question<|Assistant|><think>reasoning</think>

Compiled `flex_attention` wrapper with Torchtitan-style inductor options.

An error raised when the user receives a 403 Forbidden error when trying to create a bucket. This can occur if the bucket already exists and belo

Client for interacting with Google Maps APIs

Renderer for Kimi K2.5 with thinking disabled. Uses `<think></think>` prefill instead of `<think>` to disable thinking mode. Format:

Renderer for Kimi K2.5 with thinking enabled (default). Key differences from KimiK2Renderer: 1. Generation prompt prefill: Appends `<thi

Format for moonshotai/Kimi-K2-Thinking: <|im_system|>system<|im_middle|>You are Kimi, an AI assistant created by Moonshot AI.<|im_end|>

Stateful streaming parser for Kimi K2 model output. Parses tokens incrementally, yielding deltas as content becomes available. Handles UTF-8

Renderer for Llama 3 Instruct models. Format:: <|begin_of_text|><|start_header_id|>system<|end_header_id|> You are a helpful AI

An error raised when deployment of a LoRA times out. Status code: 504

Result from LocalBackend.train(). Attributes: step: The training step after this training call completed. metrics: Aggregated tra