jondurbin/airoboros-c34b-2.2.1

Model Overview

jondurbin/airoboros-c34b-2.2.1 is an experimental 34 billion parameter language model, building upon the airoboros-c34b-2.2 branch. It is primarily fine-tuned for strong instruction following, distinguishing itself from models optimized for casual chat or roleplay. The model incorporates several key updates:

Key Capabilities

• Enhanced Instruction Following: Focuses heavily on precise instruction adherence across various tasks.
• Context-Obedient QA: Trained to ignore prior knowledge and strictly use provided context for question answering, reducing hallucinations. It utilizes a specific BEGININPUT/BEGINCONTEXT/BEGININSTRUCTION format for closed-context tasks.
• Summarization: Includes training for summarizing text, using a similar delimited format as contextual QA.
• Code Generation: Capable of generating complex code based on detailed requirements, with an optional PLAINFORMAT for raw code output.
• Agent/Function Calling: Supports generating JSON or YAML for function/argument selection, similar to OpenAI's function calling.
• Chain-of-Thought Reasoning: Can provide multiple potential answers, rank them by logic, and select the most feasible one.
• reWOO-style Execution Planning: Generates systematic plans for complex instructions requiring multiple tool calls, outputting a sequence of function calls and evidence references.

Training Details

The model was trained for 5 epochs, incorporating re-generated writing responses, longer contextual blocks, and de-censoring. The training data was largely synthetic, generated by the airoboros tool, primarily using OpenAI API calls to GPT-4.

Usage Notes

Users should implement stopping criteria (e.g., on "USER:") for chat scenarios. The model's license is ambiguous due to its reliance on OpenAI API-generated data, suggesting caution for commercial use.