Specialize / Private Fine-Tuning

Hebbian specializes models only when the evidence says it is worth it.

We measure first. If a general model cannot meet the required judgment, we design the corpus, benchmark, and training with explicit limits.

benchmark firstprivate corpuspublic paper
Paper #001published benchmark
Modelmeasured gap
Trainingevidence first

Evaluation protocol

Train with evidence.

Hebbian compares models, reviews answers, and decides whether private training makes operating sense.

01Baseline

What the general model does today.

02Criterion

What quality, latency, or consistency is missing.

03Gap

What difference would justify training.

Evidence before training

Benchmark before corpus.

That keeps training from becoming enthusiasm and focuses the work on visible improvement.

  • Documented comparative benchmark.
  • Private corpus with usage criteria.
  • Honest reading of limits and results.

Lab method

Lab method.

01Measure

Open a baseline with clear prompts and categories.

02Curate

Prepare examples with intent and quality control.

03Train

Run training only when the gap justifies it.

04Read

Document findings, limits, and the next decision.

Success criteria

Proven improvement.

Score

Answers improve in relevant categories.

Latency

The system responds within the useful margin.

Judgment

Limits remain visible for the next decision.

What remains

Laboratory judgment.

Fine-tuning is treated as a business decision: if the gap does not show up in evidence, training is not forced.

01Baseline

The current model is measured before touching corpus or infrastructure.

02Honest comparison

Score, latency, loops, and limits are read together to avoid cosmetic conclusions.

03Executive decision

The result clarifies whether to train, change approach, or strengthen the benchmark.

Hebbian Research Notes

Paper #001.

The Specialize page keeps the lab visible: benchmark, findings, limits, and a documented run.

Real artifact

The first gallery documents the LoRA run from April 24, 2026.