FinTFM (synthetic-only) multiclass checkpoint: dr74-mc-s0.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-multiclass,
Apache-2.0. Code: github.com/kabartay/fintfm.
| family | FinTFM (synthetic-only). No real data. |
| local path | runs/kaggle-ts-dr74-mc-s0/dr74-mc-s0/dr74-mc-s0.pt (gitignored, never committed) |
| size | 3,598,528 bytes |
| SHA-256 | 14d5b4d561fbea6e35c7a30dc2644cb8f64f437c25d62da087ee960717cc8f4d |
| parameters | 888,090 (same architecture as the binary checkpoint, max_classes 10) |
| used by | the multiclass arm of the TabArena entry: default checkpoint of runs/local_queue/lite_mc_probe.sh, run fintfm_lp_mc_scr_bag8_1113 |
Intended use
Frozen representation for a per-leaf readout on multiclass tabular tasks of up to 10 classes and 136 features (FinTFM-LinProbe with partition bags). Research and benchmarking.
Training data
Synthetic prior only: --max-classes 10 --p-financial 0.5 --scm-legacy --permute-classes,
so half the tasks come from the financial prior and half from the generic structural-causal-model
prior. 6,000 steps at a global batch of 8 = 48,000 tasks, seed 0, Kaggle T4x2 with DDP,
fp16, split attention (scripts/kaggle/ts_dr74_mc_s0/; pre-registered in §185; fintfm pinned
at a901cdc per §185, package version 0.7.1 per the stored provenance; the stored commit
field is empty). Recipe copy: runs/kaggle-ts-dr74-mc-s0/dr74-mc-s0/recipe.yaml. No real data.
Evaluation (all MEASURED)
- Proxy suite, 6 of 8 multiclass tasks scored (
letterandmfeat-factorsexceed the caps): 994 Elo against 808 for the earlier multiclass checkpoint; 6 task wins, 0 losses (§186). - TabArena-Lite multiclass, entry run
fintfm_lp_mc_scr_bag8_1113: 874 Elo over 8 datasets (§227, §243); four bags gave 865 (§224). Entry overall: 931.9 Elo (§235).
Limitations
- Raw-only control (§243). Without the representation (same configuration,
no_rep) the entry scores 829 Elo against 874; the representation is better on 7 of 8 datasets (sign test p = 0.07), median log loss 15.5% lower, and worse onhiva_agnosticby 2.0%. Unlike binary, the multiclass Elo is carried by the representation, on eight datasets and one seed. - Eight datasets, one repeat, one fold; the Elo interval for the control is +151 / -344.
- Digit-feature tables (
mfeat-*) stay far behind linear and boosted models (§186). - More than 10 classes or 136 features is out of range.
LICENCE: Apache-2.0 (owner decision 2026-11-16, matching the code and the public binary checkpoint)