recall-honcho-8b — a local Honcho deriver (explicit conclusion extraction)

A Qwen/Qwen3-8B fine-tune specialised for the explicit conclusion-derivation step of Honcho: given a target peer and their chat turns, emit atomic, self-contained, correctly-attributed facts as Honcho-schema JSON. Independent, self-hosted re-creation of the role Plastic Labs' (closed) Neuromancer XR plays in production Honcho. Not affiliated with Plastic Labs.

Results

Held-out validation (298 examples, 4% split), bf16 LoRA, 2 epochs:

Metric	Value
Eval loss	0.201
Eval token accuracy	93.8%
Train loss (final)	0.197
Loss curve	2.03 → 0.13 over 896 steps

The model reliably reproduces the exact {"explicit":[{"content":...}]} schema with correct attribution, absolute dates, and atomic facts. Worked example (held-out):

Input  : message about buying a secondhand Eames lounge chair
Output : {"explicit": [
  {"content": "dmitri bought a secondhand Eames lounge chair from a Facebook listing in Williamsburg"},
  {"content": "dmitri paid $900 for the secondhand Eames lounge chair"},
  {"content": "dmitri rented a Zipcar to haul the Eames lounge chair back, costing another $80"} ]}

LoCoMo benchmark (base Qwen3-8B vs this model, through Honcho's dialectic pipeline) — to be added. Reference points from Plastic Labs' blog (different/private data, not directly comparable): base Qwen3-8B 69.6, Claude 4 Sonnet 80.0, Neuromancer XR 86.9.

Training

• Base: Qwen/Qwen3-8B
• Data: 7,160 synthetic SFT examples, gold labels distilled from Claude Opus 4.8 (frontier teacher), 15 balanced life-domains, rendered through Honcho's own prompt-builder and schema-validated against PromptRepresentation. Distilling a compact deriver from a frontier model is the core idea.
• Method: bf16 LoRA (rank 32, alpha 32), 2 epochs, lr 2e-4 cosine, completion-only loss.
• Hardware: NVIDIA RTX PRO 6000 Blackwell (sm_120), ~56 min.

Prompt / I-O contract

Input = Honcho's minimal_deriver_prompt(peer_id, messages) verbatim (non-thinking). Output = JSON validated by PromptRepresentation:

{"explicit": [{"content": "alice is training for a half-marathon scheduled for October 2026"}]}

Rules: atomic · self-contained · absolute dates · correct third-party attribution · explicit-only (no speculation) · {"explicit": []} when nothing is stated.

How to use with Honcho

Serve (vLLM/Ollama), expose via your gateway (e.g. litellm route recall-honcho-8b), then point Honcho's deriver at it in config.toml:

[deriver.model_config]
transport = "openai"
model = "recall-honcho-8b"
[deriver.model_config.overrides]
base_url = "http://<gateway>/v1"
api_key_env = "LLM_OPENAI_API_KEY"

Keep the dialectic/generation step on a larger model.

Files

• Merged bf16 model (root) — servable standalone.
• lora-adapter/ — the LoRA adapter alone (apply onto Qwen/Qwen3-8B).

Limitations

• Explicit level only (not deductive/inductive/abductive — Honcho's dreamer).
• English only; trained on privacy-safe synthetic data (frontier-distilled).
• Optimised for Honcho's exact prompt; off-format prompts may degrade output discipline.

Acknowledgement

Inspired by Plastic Labs' Honcho and Neuromancer research. Independent synthetic re-creation; not affiliated with or endorsed by Plastic Labs.