s-nlp/ToolHACE_qwen3.5_2B_span

VISIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2.3BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 20, 2026Architecture:Transformer Featherless Exclusive Cold

ToolHACE Qwen3.5 2B Span by s-nlp is a 2.3 billion parameter Qwen3.5-based model fine-tuned as a tool-calling hallucination auditor. It compares user requests, available tools, tool outputs, and assistant answers to return a JSON verdict with hallucination classes and verbatim answer spans. This model is specifically designed to audit the final answer of a tool-calling assistant for factual consistency against provided tool outputs. It excels at identifying various types of hallucinations, including answer mismatches, overgeneration, missing tools, and undergeneration.

Loading preview...

Model Overview

s-nlp/ToolHACE_qwen3.5_2B_span is a 2.3 billion parameter model built on the Qwen3.5-2B architecture. It is specifically fine-tuned to act as a tool-calling hallucination auditor, generating a JSON verdict that identifies hallucination classes and relevant answer spans. Unlike conventional token classification models, it produces a generative JSON output.

Key Capabilities

  • Hallucination Detection: Audits the final answer of a tool-calling assistant by comparing it against the user request, available tools, and tool outputs.
  • Generative JSON Output: Provides a structured JSON verdict detailing hallucination types and exact text spans from the assistant's answer.
  • Identifies Specific Error Types: Detects answer_mismatch, overgeneration, missing_tool, and undergeneration (where the span is null).
  • High Precision: Achieves a response precision of 0.9820 and a character-level span F1 of 0.8773 in test evaluations.

Intended Use Cases

  • Post-execution Auditing: Ideal for auditing the final answers of tool-calling assistants to ensure factual consistency with tool outputs.
  • Quality Assurance: Helps in identifying and categorizing different types of hallucinations in AI-generated responses within tool-use scenarios.

Limitations

  • Primarily trained on English tool-calling dialogues with generally one error class per row.
  • Struggles with undergeneration detection, showing a lower recall (0.0423) for this category.
  • Answers exceeding 4,000 characters are truncated due to prompt contract limits.

Finetunes
1 models