Language structure
Text compression and basic sequential dependencies.
CodeSOTA field guide · base models · 2026
A practical benchmark ladder for models from 1M to 7B parameters. Each test enters when it produces a signal—after performance separates from chance, but before saturation.
The model is no longer behaving randomly.
The score has not yet saturated.
Checkpoints differ by more than measurement error.
Scoring does not require instruction tuning.
Text compression and basic sequential dependencies.
The first external language signal, but still no knowledge KPI.
The transition from linguistic form to lexical knowledge.
The most informative range for classic small base models.
Harder distractors take over from simple knowledge tests.
Broader knowledge and common sense; report MMLU by category.
Math and code only now become stable parts of the suite.
General tests give way to harder domain evaluations.
Measure prose, news, dialogue, code, and technical documents separately. BPB remains the primary metric for 1–50M models.
Report the mean margin and soft score σ(m) alongside accuracy. A small model can improve its distribution before changing the binary decision.
For 50–150M models, use target-token NLL, MRR, and top-k. Exact match becomes a primary metric only after it stops being nearly zero.
Also report the probability of the correct answer and its margin over the strongest distractor. HellaSwag without normalization favors shorter completions.
The Pythia family shows how useful ranges overlap. SciQ produces a strong signal early and approaches its ceiling; ARC-Challenge starts late and improves slowly. ARC-Easy and LAMBADA bridge the middle of the scale.
| Model | SciQ | ARC-Easy | LAMBADA | ARC-Challenge |
|---|---|---|---|---|
| 70M | 60.1% | 37.4% | 18.5% | 18.1% |
| 160M | 74.1% | 43.5% | 32.8% | 18.8% |
| 410M | 81.1% | 52.1% | 51.6% | 21.3% |
| 1B | 84.0% | 56.9% | 56.2% | 24.4% |
| 2.8B | 88.2% | 64.4% | 64.7% | 29.5% |
| 6.9B | 89.7% | 67.3% | 67.3% | 31.3% |
The Pythia reference values illustrate dynamic range; they do not constitute a new model ranking.
Accuracy is near chance, but the soft score or margin begins to move. Example: SciQ at 28% against a 25% chance baseline for a 50M byte-level model.
The score is clearly above chance, differences exceed measurement error, performance remains below 90–95%, and the prompt does not dominate the result.
The test stays in the report but leaves the aggregate. This applies, for example, to SciQ at several billion parameters or saturated BLiMP categories.
IFEval, AlpacaEval, MT-Bench, format adherence, and tool use require instruction tuning. They should not lower the assessment of a raw pretraining checkpoint.