dman1011/recall-honcho-8b

TEXT GENERATIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:May 31, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

dman1011/recall-honcho-8b is an 8 billion parameter Qwen3-8B fine-tune, specifically optimized for explicit conclusion derivation within the Honcho framework. This model excels at extracting atomic, self-contained, and correctly attributed facts from chat turns, outputting them in a precise JSON schema. It achieves a 93.8% evaluation token accuracy on held-out validation, making it suitable for structured fact extraction in conversational AI applications.

Loading preview...

Overview

dman1011/recall-honcho-8b is an 8 billion parameter model fine-tuned from Qwen/Qwen3-8B, designed for the explicit conclusion-derivation step of the Honcho framework. It functions as an independent, self-hosted re-creation of Plastic Labs' Neuromancer XR, focusing on extracting atomic, self-contained, and correctly attributed facts from peer chat turns into a Honcho-schema JSON format.

Key Capabilities

  • Explicit Conclusion Derivation: Specializes in identifying and structuring explicit facts from conversational data.
  • Schema Adherence: Reliably reproduces the exact {"explicit":[{"content":...}]} JSON schema, ensuring correct attribution, absolute dates, and atomic facts.
  • High Accuracy: Achieved an evaluation token accuracy of 93.8% on a held-out validation set, with an eval loss of 0.201.
  • Distilled Intelligence: Trained on 7,160 synthetic SFT examples, with gold labels distilled from Claude Opus 4.8, making it a compact deriver from a frontier model.

Training and Usage

The model was trained using bf16 LoRA (rank 32, alpha 32) over 2 epochs, with a learning rate of 2e-4 cosine. It processes Honcho's minimal_deriver_prompt(peer_id, messages) as input and outputs JSON validated by PromptRepresentation. It is designed to be served via vLLM/Ollama and integrated into Honcho's deriver configuration.

Limitations

  • Explicit Level Only: Focuses solely on explicit facts; does not perform deductive, inductive, or abductive reasoning.
  • English Only: Trained exclusively on privacy-safe synthetic English data.
  • Prompt Sensitivity: Optimized for Honcho's exact prompt format; off-format prompts may lead to degraded output discipline.