Measure the rung.
Keep the evidence.

A private GPU scorer for deterministic language-model benchmarks. Same pair IDs, same tokenizer boundaries, same fp32 protocol—ready for controlled small-model comparisons.

Run a micro battery

Paste a Hugging Face model.

The RTX 3090 loads the model, scores up to 2,000 provisional Polish agreement pairs, then returns continuous pair probabilities. First run includes model download.

The Slayer mission

Make Polish model scaling predictable.

We are not starting with a giant model and hoping the measurements catch up. We are building an instrument that tells us what the next rung can teach us.

RungScalePrimary signalDecision
018MBPB · morphology · agreementFoundations
0216MBPB · minimal pairsLanguage structure
0332MPairs · short compositionData mix
0464MStructured language · reasoning-liteCapability separation
05125MMC correct-choice BPBUseful comparison
06250MRecall · composition · basic codeScaling check
07500MMulti-step tasksPrediction update
08–121B → 20BReasoning · tools · long contextStop or scale
At every rung

Measure loss, morphology, pairs, composition and useful capability only when the scale can separate it.

The discipline

Train, measure, compare, update the curve, predict the next scale, then spend the next order of magnitude of compute.

The artifact

The valuable output is the curve connecting 8M to 20B: what moved it, what saturated and what was worth the compute.

Loaded modelKoliber v1.1
Parameters126.0M
HardwareRTX 3090
Protocolfp32 · MPS/CUDA

Measured Polish micro

2,000 identical provisional agreement pairs. Pair probability is continuous; accuracy is shown only as a diagnostic. BPB is not included because the frozen validation corpus is not yet available.

ModelSentence pair probabilityCritical-region probabilityTime
Koliber v1.10.90830.8913111.3 s
Pollock 1.40.59140.5718100.4 s

What this service measures

Continuous

Sentence and critical-region pair probabilities, with raw log margins retained.

Deterministic

One worker, serialized requests, fixed protocol and no generation sampling.

Auditable

Revision, parameter count, tokenizer hash, device and runtime are returned.

Private

Loopback-only by default. Reach it through SSH forwarding or a private authenticated proxy.

The workflow

Train 32M → measure it → compare against 16M → update scaling curves → predict 64M → train 64M → check the prediction.

If a larger run misses its predicted loss or capability profile, stop and understand why. The benchmark is a measurement instrument with different sensitivity ranges, not a static collection of standard scores.

API

Run the micro battery from a model id or URL, or send your own one to 2,000 pair records. The service serializes GPU work.

curl -X POST https://eval.fabryka.ai/v1/benchmark/micro \
  -H 'content-type: application/json' \
  -d '{"model":"org/model","limit":2000}'

curl -X POST https://eval.fabryka.ai/v1/score/pairs \
  -H 'content-type: application/json' \
  --data-binary @pairs-request.json