r"""Arguments pertaining to what data we are going to input our model for training and evaluation.

r"""Dataset attributes.

r"""General class for chat models. Backed by huggingface or vllm engines. Supports both sync and async methods. Sync methods: chat(), stream_

r"""Arguments pertaining to which techniques we are going to fine-tuning with.

r"""A callback for saving the processor.

r"""Arguments pertaining to which model/config/tokenizer we are going to fine-tune or infer. The class on the most right will be displayed first.

r"""Data collator that supports VLMs. Features should contain input_ids, attention_mask, labels, and optionally contain images, videos and audios

r"""A class to manage all the gradio components in Web UI.

r"""Compute and store the average and current value.

r"""Inherits Seq2SeqTrainer to compute generative metrics such as BLEU and ROUGE.

r"""A dummy optimizer used for the GaLore or APOLLO algorithm.

r"""A general engine to control the behaviors of Web UI.

items: list[dict] with fields: - input_ids: LongTensor[L] - labels: LongTensor[L] (question tokens = -100; CoT tokens = 0/1)

Reads the torch-saved list of dict samples (exact token-level alignment).

r"""Data collator for pairwise data.

r"""Compute reward accuracy and support `batch_eval_metrics`.

r"""Compute accuracy and support `batch_eval_metrics`.

r"""Compute text similarity scores and support `batch_eval_metrics`. Wraps the tokenizer into metric functions, used in CustomSeq2SeqTrainer.

r"""Inherit PPOTrainer.

r"""Inherit Trainer for custom optimizer.

r"""Default tool using template.

r"""GLM-4-MOE tool using template.

r"""GLM-4 tool using template.

r"""Data collator for KTO data.

r"""Ling v2 tool using template.

r"""Llama 3.x tool using template with `tools_in_user_message=False`. Reference: https://www.llama.com/docs/model-cards-and-prompt-formats/llama3

r"""A callback for logging training and evaluation status.

r"""Mistral v0.3 tool using template.

r"""Data collator for pairwise data.

r"""Inherits Trainer to compute pairwise loss.

r"""A callback for converting the PiSSA adapter to a normal one.

r"""Qwen 2.5 tool using template.

r"""A callback for reporting training status to external logger.

r"""A class to manage the running status of the trainers.

r"""Data collator for 4d attention mask.

Inference engine for SGLang models. This class wraps the SGLang engine to provide a consistent interface for text generation that matches LLa

r"""Seed tool using template.

ratio_encoding: "numeric" | "special" | "none" - numeric: * qwen -> append "<|eot_id|>{ratio:.1f}<|eot_id|>" to user (only rati

通过 /v1/completions 发送与本地完全相同的字符串 prompt。 一次请求用 n>1 返回多采样，便于做 best-of。

Weighted Focal Loss with ignore_index masking.

Weighted Focal Loss with proper ignore_index masking.

r"""Arguments pertaining to the APOLLO algorithm.

r"""Arguments pertaining to the BAdam optimizer.

r"""Base class for inference engine of chat models. Must implements async methods: chat(), stream_chat() and get_scores().

r"""Arguments pertaining to the model.