altslate/certo-decision-v2.2

Certo v2.2 (Ettin-encoder-1B) — UNDERTRAINED checkpoint, for continuation

Honest status: this is an under-trained, one-epoch checkpoint that currently performs worse than the 150M base model (certo v2.1). It is published so training can be continued, not for use as-is.

What it is

A generic decision model on the Ettin-encoder-1B backbone (a ~1B-parameter ModernBERT-architecture encoder). Non-generative: scores supplied runtime options and returns a calibrated distribution. Trained on the representative real+synthetic corpus altslate/certo-decisions-v2 (v2.1, 971k). Temperature-calibrated (T=1.33).

Why it underperforms (be clear about this)

It ran only 1 epoch (a 1B model on 4×32 GB is 6× slower than base; 3 epochs ≈ 2.7 days). It did not converge: final training loss 0.458 vs the base model's **0.34** after 3 epochs. Under-training shows up as both lower accuracy and poor calibration.

JevBench public tasks (official certo_local adapter, calibrated):

	easy	original	hard
v2.2 (Ettin-1B, 1 ep)	0.312	0.319	0.297
ECE	0.33	0.33	0.42
v2.1 (base, 3 ep) for reference	0.583	0.389	0.297

So on this budget the 1B backbone is not an improvement — the result is dominated by under-training, not by model size.

Intended use: continue training

# warm-start and train more epochs (see AltSlate-Labs/certo train_v2.py --init)
python train_v2.py <corpus> --backbone jhu-clsp/ettin-encoder-1b --init <this_checkpoint> \
    --epochs 2 --bs 16 --grad_checkpoint

Load for inference (works, but expect the weak numbers above):

from certo import DecisionModel
m = DecisionModel.load("altslate/certo-decision-v2.2", revision="main")

Code/recipe: github.com/AltSlate-Labs/certo.