Qwen3.5 chat template expects arguments as a mapping for |items.

Train the model using supervised fine-tuning. For ServerlessBackend, this serializes trajectories to a JSONL file, uploads it to W&B

Train the model using supervised fine-tuning. Args: model: The trainable model to fine-tune trajectories: Iterable of

Format a trajectory into a readable log string.

Reconstruct a Renderer from its name and model name. Called by pickle to deserialize Renderer instances. Uses cached tokenizer/image_processor

Execute a tool with given arguments. Args: tool_name: Name of the tool to call arguments: Arguments to pass to the to

Apply a tool with given arguments. Args: tool_name: Name of the tool to call arguments: Arguments to pass to the tool

Execute a tool via the API server. Args: tool_name: Name of the tool to call arguments: Arguments to pass to the tool

Build generation prompt, prepending system message if configured.

Build generation prompt with <think></think> prefill to disable thinking.

Build generation prompt with <think> prefill to trigger thinking mode. Does NOT add <think> when the previous message is a tool response,

Build supervised example, prepending system message if configured.

Build tokens and per-token weights for supervised fine-tuning. This function returns a list of examples in the form of tuples, where

Build tokens and per-token weights for supervised fine-tuning. Since Kimi K2 renderer does not satisfy the extension property, this method is

Calculate correlation and RMSE metrics for each split in the dataset. Args: df: DataFrame with log_score, predictions and split colu

Return compression ratio (original_size / new_size).

Consume a chat completion stream and build a complete ChatCompletion object. This function processes a stream of ChatCompletionChunks, constructi

Create conversation prefix with tools in Harmony format. Returns a list of messages to prepend to conversations: 1. If tools present:

Create system messages with TypeScript-style tool specifications. Per the HuggingFace chat template, Kimi K2.5 uses TypeScript-style tool

Create system message with Qwen3.5 tool specifications. Qwen3.5 uses a different tool declaration format from Qwen3, with XML-based f

Create system message with DeepSeek V3.1 tool specifications. DeepSeek V3.1 tool calling requires tools to be described in the system message

Create system message with Qwen3 tool specifications. Qwen3 uses XML `<tools>` tags containing JSON tool definitions in OpenAI format,

Create system messages with Kimi K2 tool specifications. Per the HuggingFace chat template, Kimi K2 places the tool_declare message B

Put the worker to sleep, offloading both weights and KV cache. Args: level: The sleep level: - 1: offload KV cache to CP

Wake up the worker from sleep, restoring offloaded weights and KV cache.

Check if fixtures are available.

Format message content as a string, preserving all part types. Unlike get_text_content which only extracts text parts, this formats all conte

Compute self attention with compiled flex kernels. Args: query: `[S, B, Hq, D]` key: `[S, B, Hkv, D]` val

Create a Resource from a dictionary.

Create a GeneratedScenario from a dictionary.

Create a GeneratedScenarioCollection from a JSON string.

Generate scenarios for MCP tools. Args: tools: List of Tool objects or list of tool dictionaries resources: Optional list of

Get a model-specific mutliplier for the LR, when training with LoRA. Given two models A and B, and learning rate LR_A that's known to be opti

Get the number of parameters in the LoRA adapter.

Return a list of all registered custom renderer names.

Return a list of all registered custom tokenizer names.