FinTFM (synthetic-only) pure-regression checkpoint: v4-binned48_3-control.pt

LICENCE: Apache-2.0 (owner decision 2026-11-16, matching the code and the public binary checkpoint)

Status: public since 2026-11-17 at kabartay/fintfm-regression, Apache-2.0. Code: github.com/kabartay/fintfm.

The file name says "control" because it began as the matched binned-head control for the pinball-quantile experiment (§148). It is the checkpoint the entry uses for regression.

family FinTFM (synthetic-only). No real data.
local path runs/task48-3-kaggle/v4-binned48_3-control.pt (gitignored, never committed)
size 3,595,783 bytes
SHA-256 31d5b3ef49e89bbe405fc85ab1d746a9fd6c3a1c5282e1770111902cc4af526c
parameters 888,090 (binned head, max_classes 10)
used by the regression arm of the TabArena entry, run fintfm_lp_reg_rawpart4_pure_1114; its queue entry reg-part-bag4 sets REG_CKPT to this path (the default REG_CKPT of lite_reg_probe.sh is the older mixed v4-regression.pt, which the entry does not use)

Intended use

Frozen representation for a per-leaf ridge readout on regression tasks (FinTFM-LinProbe with four partition bags). Research and benchmarking.

Training data

Synthetic prior only: --p-regression 1.0 (every task a synthetic regression task), binned head, --max-classes 10, 6,000 steps x batch 8 = 48,000 tasks, seed 0, Kaggle GPU, fintfm==0.5.6 from PyPI (scripts/kaggle/task_48_3_binned_control/; §148). The checkpoint stores no commit (no provenance field); the kernel script and the pinned package version are the record. No real data.

Evaluation (all MEASURED, TabArena-Lite regression, repeat 0 fold 0, CPU, 13 datasets)

  • First use in place of the mixed checkpoint, one bag: 582 Elo against 578, RMSE lower on 8 of 13 datasets (§227).
  • Entry run fintfm_lp_reg_rawpart4_pure_1114: 689 Elo, better than the one-bag run on 12 of 13 datasets (sign test p < 0.01) (§235).

Limitations

  • Raw-only control (§240). Without the representation (no_rep, effectively four bagged trees with a ridge per leaf) the entry scores 537 Elo against 689; the representation is better on 12 of 13 datasets (p = 0.003), median RMSE 11.4% lower, largest on airfoil_self_noise where raw only is 88.6% worse. Food_Delivery_Time, QSAR_fish_toxicity and QSAR-TID-11 differ by under 1%.
  • A 480,000-task pure-regression checkpoint did not lift regression (§233); the 48,000-task one is what the entry uses. One seed per arm, one fold.

LICENCE: Apache-2.0 (owner decision 2026-11-16, matching the code and the public binary checkpoint)

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support