Skip to content
Models · The PredictLM family

Two checkpoints. One interface. Open weights.

PredictLM is a family of open-weight tabular foundation models. Same architecture family, same calibrated-uncertainty head, two size tiers. Both Apache-2.0.

Ship recipe · Duo + TTT

0.751 classification accuracy. 0.609 regression R². Locked 25-dataset OpenML benchmark.

The recommended way to use PredictLM is the Duo + TTT recipe: load both models, run ~15 inner Adam steps of self-supervised fine-tuning on the user's in-context examples (test-time training), then average their softmax predictions at w = 0.40. +7.8 percentage points of classification accuracy and +7.3 pp of regression R² over zero-tuning. Implemented in one Python call.

Classification

0.751

mean accuracy (n=10 datasets)

vs 0.673 Mini-v1 alone (+7.8 pp) · 0.685 Base alone (+6.6 pp)

Regression

0.609

mean R² (n=10 datasets)

vs 0.536 Mini-v1 alone (+7.3 pp) · 0.589 Base alone (+2.0 pp)

19 of 20 evaluated datasets improved with TTT versus zero-tuning; no dataset regressed by more than 0.006. Single-model alternatives: Mini-v1 + TTT alone gets 0.742 / 0.595 (one model loaded); Base + TTT alone gets 0.748 / 0.608.

Recommended · Edge / CPU

PredictLM‑Mini

13M parameters · 54 MB · Apache-2.0

The smallest open-weight tabular FM with calibrated uncertainty. Distilled from PredictLM-Base. Statistically tied with Base on classification accuracy and within ~4 percentage points (pp) R² on regression. Runs on any laptop.

What it does

  • Runs on a laptop — CPU only, no GPU required
  • Returns full prediction confidence intervals, not just guesses
  • ~95% of Base's accuracy at half the size

Evaluation

  • Classification: mean accuracy = 0.673 (25-dataset OpenML)
  • Regression: mean R² = 0.536
  • Statistically tied with Base on classification
  • Trained in 3.3 hours for ~$1.30 of cloud compute
View on Hugging Face

PredictLM‑Base

26M parameters · 105 MB · Apache-2.0

The best-accuracy model in the family. Teacher for Mini and the architecture of record for our published evaluations. Trends ahead of XGBoost on regression; CI within sampling noise.

What it does

  • Highest-accuracy model in the family
  • Returns full prediction confidence intervals, not just guesses
  • Reference architecture for the PredictLM family

Evaluation

  • Classification: mean accuracy = 0.685
  • Regression: mean R² = 0.589
  • With Mini + test-time training: 0.751 / 0.609
View on Hugging Face
Side-by-side

Mini vs Base

SpecMiniBase
Parameters13M26M
Checkpoint size54 MB105 MB
LicenseApache-2.0Apache-2.0
Best forCPU / edgeHighest accuracy
Classification (alone)0.6730.685
Regression R² (alone)0.5360.589
With test-time training0.742 / 0.5950.748 / 0.608
Combined (Duo + TTT recipe)0.751 cls / 0.609 reg

Single-model numbers are point estimates on the same locked 25-dataset OpenML benchmark. The Duo + TTT recipe uses both checkpoints together.

How to use

Same model. Two ways to call it.

Python · one call, ship recipe by default
pip install predictlm

from predictlm import PredictLM

model = PredictLM.from_pretrained("zerooneresearch/predictlm-mini-13m")

# Just .fit().predict() — the package silently downloads the partner
# checkpoint and runs the published Duo + TTT ensemble under the hood.
# Returns the 0.751 cls / 0.609 reg result on the locked OpenML eval.
preds = model.fit(X_train, y_train).predict(X_test)

# Single-model fast path (no Duo, no TTT) — pass auto_duo=False:
# model = PredictLM.from_pretrained(..., auto_duo=False)
# preds = model.fit(X_train, y_train).predict(X_test)   # ~0.673 cls
MCP tool (Claude / Cursor / Continue)
pip install predictlm-mcp

# in ~/Library/Application Support/Claude/claude_desktop_config.json:
{"mcpServers": {"predictlm": {
  "command": "predictlm-mcp",
  "args": ["--model", "predictlm-mini-13m"]
}}}