Train Local, Ship Portable

Fine-tune retrieval models locally, publish portable GGUF artifacts, benchmark them against the shipped baseline, and use them in GNO with HF-backed custom presets.

Key Benefits

Local MLX LoRA training on Apple Silicon
Automatic checkpoint selection
Portable GGUF export
Real benchmark-based promotion
Custom preset install snippets

Example Commands

gno models use slim-tuned gno models pull --gen gno query 'ECONNREFUSED 127.0.0.1:5432' --thorough

Get Started

Ready to try Fine-Tuned Models?

Quick Start Guide CLI Reference

Why This Matters

Fine-tuning is only useful if the resulting model can be exported, benchmarked, and installed cleanly. GNO’s fine-tuning workflow is built around that full loop:

train locally
select the best checkpoint
export a portable GGUF
benchmark the exported artifact
publish the promoted artifact
install it in GNO with an hf: preset

Current Promoted Model

The current promoted slim retrieval model is slim-retrieval-v1, produced from auto-entity-lock-default-mix-lr95.

repeated benchmark median nDCG@10: 0.925
repeated benchmark median schema success: 1.0
repeated benchmark median p95: 4775.99ms
HF repo: guiltylemon/gno-expansion-slim-retrieval-v1

Local Training, Portable Artifact

The current local training backend is MLX LoRA on Apple Silicon. That is a training implementation detail, not a deployment requirement.

The deployable artifact is a GGUF file that GNO can load via hf: or file: URI.

models:
  activePreset: slim-tuned
  presets:
    - id: slim-tuned
      name: GNO Slim Retrieval v1
      embed: hf:gpustack/bge-m3-GGUF/bge-m3-Q4_K_M.gguf
      rerank: hf:ggml-org/Qwen3-Reranker-0.6B-Q8_0-GGUF/qwen3-reranker-0.6b-q8_0.gguf
      gen: hf:guiltylemon/gno-expansion-slim-retrieval-v1/gno-expansion-auto-entity-lock-default-mix-lr95-f16.gguf

Use file: only for private or unpublished models.

Promotion Flow

One command drives the full promotion path:

bun run research:finetune:finalize slim-retrieval-v1 auto-entity-lock-default-mix-lr95

That command:

materializes the canonical release bundle
writes a public model card and install snippet
stages the HF upload bundle

For the full training-to-promotion flow:

bun run research:finetune:promote <run>

Why Benchmark The Exported Model

Training loss is not enough.

The model that looks best during training can still perform worse after export on real retrieval tasks. GNO promotes based on exported-model benchmark results, not loss alone.