jondurbin/airoboros-34b-3.2

TEXT GENERATIONConcurrency Cost:2Model Size:34BQuant:FP8Ctx Length:32kPublished:Mar 4, 2024License:yi-licenseArchitecture:Transformer0.0K Cold

jondurbin/airoboros-34b-3.2 is a 34 billion parameter experimental language model fine-tuned by jondurbin, based on the Yi-34B-200K architecture. It was trained primarily on synthetic data generated by the Airoboros project, augmented with several third-party datasets, and is optimized for multi-turn conversations, context-obedient question answering, summarization, complex coding tasks, and agentic function calling. While its base model supports a 200K context window, this fine-tuned version is effective up to 8K tokens.

Loading preview...

Model Overview

jondurbin/airoboros-34b-3.2 is an experimental 34 billion parameter language model, fine-tuned by jondurbin using the Yi-34B-200K as its base. It leverages a custom dataset, airoboros-3.2, which includes multi-turn data and "toxic" instructions, alongside several third-party datasets like bluemoon-fandom-1-1-rp-cleaned, glaive-function-calling-v2, and Vezora/Tested-22k-Python-Alpaca.

Key Capabilities

  • Context-Obedient Question Answering: Specifically tuned to prioritize provided context over its internal knowledge, reducing hallucinations. It uses a verbose BEGININPUT/BEGINCONTEXT/BEGININSTRUCTION format for closed-context prompts.
  • Summarization: Includes training for summarization tasks, using a similar structured input format.
  • Complex Coding: Capable of generating code for intricate requirements, including multi-threaded TCP servers, FastAPI webservers, and can output plain code using the PLAINFORMAT instruction.
  • Agent/Function Calling: Supports generating JSON or YAML for function calls based on user input, similar to OpenAI's function calling.
  • Chain-of-Thought Reasoning: Can provide multiple potential answers to a problem, rank them by logical soundness, and select the most feasible one.
  • reWOO-style Execution Planning: Generates systematic plans for complex instructions requiring multiple tool calls, outputting a sequence of actions and evidence references.

Important Considerations

  • Context Length: While the base Yi-34B-200K model supports 200K context, this fine-tuned version was trained with an 8K token context size, and performance beyond this limit may be unreliable.
  • Prompt Format: Uses the Llama-2 chat format for standard interactions, with specific structured formats for advanced features like context-obedient QA and summarization.
  • Licensing: The model's license is intentionally ambiguous due to the use of OpenAI API-generated data for training. Commercial use is advised against due to potential conflicts with OpenAI's ToS.

Finetunes
1 models