A class that manages prompt templates and keeps all conversation history.

This should contains everything needed to continue the conversation. For example, the history of the conversation, the current task (success/fail

Configuration class for evaluation settings

Text generation configuration

Data structure for a single collected sample Attributes: task_id: Unique identifier for the task task_content: The original

Compressed Linear Layer.

Dataset-specific configuration

Distributed training configuration

Logging configuration

Model-related configuration

Data storage configuration

Fake file-like stream object that redirects writes to a logger instance.

Preprocess conversations and attach plan/hidden_state.

Data collector specialized for ALFWorld household task planning This collector processes ALFWorld trajectory data and generates high-level

自适应数值范围投影层

Adaptive numerical range projection layer

Extract and normalize answers from model outputs

Group-wise quantization.

Low‑rank adapter: produce rank‑r (A,B) conditioned on a context vector.

Collate examples for supervised fine-tuning with plan data.

Collator for supervised dataset.

Manages data storage in multiple formats Supports: - HuggingFace Dataset format - Parquet files with automatic sharding - Distri

Handles distributed data collection across multiple GPUs/nodes

Data class for storing evaluation results

Load hidden_state/plan data from HF datasets.

Dataset class for MMLU data

Dataset class for MMLU data

Main evaluation class

Data collector specialized for mathematical reasoning tasks This collector processes the MATH dataset and generates high-level solution plan

Handles model loading and generation with robust error handling

Monitor the number of calls to each model.

Responsible for saving extra model states (e.g., MHA weights). Rank 0 check is included.

Separator styles.

Generic dataset wrapper for task data

The model adapter for jondurbin/airoboros-*

Alfworld task instance.

The model adapter for Alpaca

The model adapter for BAAI/Aquila Now supports: - BAAI/AquilaChat-7B - BAAI/AquilaChat2-7B - BAAI/AquilaChat2-34B

The model adapter for Azure OpenAI

The model adapter for BGE (e.g., BAAI/bge-large-en-v1.5)

Model adapter for jondurbin/bagel-* models

The model adapter for Baichuan models (e.g., baichuan-inc/Baichuan-7B)

The model adapter for project-baize/baize-v2-7b

The model adapter for Bard

Base class for implementing specific data collection pipelines Subclasses should implement: - load_dataset_for_task() - create_promp

The base and the default model adapter.

The model adapter for Neutralzz/BiLLa-7B-SFT