# wikitext2_gpt2_bae A **companion** bank for [`corrected-lds`](https://github.com/simonvary/corrected-lds), built on EleutherAI **bergson**'s retrain harness rather than our own `pools/runner.py`. > **Read this first.** This bank reproduces the **Bae et al. 2024** GPT-2/WikiText-2 > setup as packaged in bergson's `examples/replicate_bae_approx_unrolling_source/` — > **not** the experiment in bergson's own paper (arXiv:2606.11660, Table 2), whose > recipe and published EK-FAC LDS (0.318) differ. > > **It reproduces the published number.** On the default `pool.npz` (eval-mode readout), > bergson's EK-FAC scores **0.4562** over Bae's five seeds, against the example's > **0.468 ± 0.015**. The pool this bank shipped before (0.3506, a "24 SE gap") was read > with **dropout active** — that was the whole discrepancy. The same models' dropout > readout is kept, explicitly labelled, as `pool_dropout_readout.npz`. > > **2026-10-25: grown to 400 masks x 20 seeds** (from 100 x 10). The first 100 masks and > seeds 1004-1013 are the previous release's cells bit for bit; every pool_tda now covers > all 400 masks and equals the previous release on masks 0-99. Numbers on > the new bank are in "Results on the 400 x 20 bank"; the tables computed on 100 x 10 are > kept and labelled as such. ## Shape ``` wikitext2_gpt2_bae/ ├── pool.npz DEFAULT — eval-mode readout (dropout off) │ R=20 seeds x M=400 masks x Q=481 queries, n_train=4,656 ├── pool_dropout_readout.npz the SAME 8,000 models, read with dropout ON │ (bergson's in-loop readout; stamped readout="dropout") ├── tauq_ekfac_bergson.npz bergson's EK-FAC, ONE full-data model (seed 1004), unprojected ├── tauq_ekfac_e10.npz our EK-FAC (kronfluence), 10 full-set runs, seeds 30000-30009 ├── tauq_ekfac_e5.npz the first 5 of those runs (seeds 30000-30004) ├── pool_tda_ekfac.npz the 10 runs, replica-resolved (R_tda = 10; E=5 = the first 5) ├── tauq_source_L3_e1.npz SOURCE (L=3, Adam variant), ONE model: bergson's scorer at │ the pin + the #511 fix, on our seed-30000 run ├── tauq_tracin_e10.npz our TracInCP over the same 10 runs x 6 half-epoch ckpts: │ the 48 transformer Linears, a fresh 16,384-dim projection │ per checkpoint ├── pool_tda_tracin.npz the 10 runs, replica-resolved (R_tda = 10) ├── tauq_trak5_e{2,4,10}.npz our TRAK, Eq. 15 units of 5: E = 2, 4, 10 (10/20/50 members) ├── tauq_trak1_e{10,20,50}.npz our TRAK, single members averaged: 10, 20, 50 members ├── pool_tda_trak5.npz the 50 members as 10 Eq. 15 units, replica-resolved └── pool_tda_trak1.npz the 50 members one by one, replica-resolved (R_tda = 50) ``` Our EK-FAC and TracIn both trace the 48 transformer Linear layers (the tied `lm_head`/ `wte`, `wpe` and LayerNorms excluded), as kronfluence and Bae et al.'s IF do. SOURCE is bergson's implementation (the pinned commit plus upstream fix #511, which restores the EK-FAC inverse in the Adam variant; without it SOURCE scored 0.13-0.17 here) at Bae et al.'s settings: 3 segments over the 6 half-epoch checkpoints, the Adam preconditioner, damping 1e-8, on our seed-30000 run (disjoint from the ground-truth seeds). Our TRAK follows the protocol of this repo's qnli bank: members trained on their OWN random halves (seeds 10000-10049) -- here by bergson, each half saved as a dataset of its own (873 steps), not by weight-zeroing -- one featurize pass giving both `trak5` (Eq. 15 over units of 5) and `trak1`. It traces EVERY parameter, as qnli's TRAK does: on five members the 48 Linears scored 0.120 against 0.169 for all parameters (TRAK's kernel inverse whitens the tied embedding that swamps TracIn's dot product). The model output is the sum of per-token log-odds (TRAK App. F.3) and `Q` the mean over tokens of `1 - p`; projection k = 2,048, the peak of a 512-4,096 sweep and the `k < n/2` guard. The members' eval-mode query loss (3.156-3.162) sits on the pool's half-data retrains (3.158). | field | value | |---|---| | model | GPT-2 (124M), fine-tuned from the base checkpoint per (mask, seed) | | data | `EleutherAI/bergson-wikitext-2-4656-chunks` @ `971da0e6da181b9e147081087f1d3fd51669024c` | | recipe | AdamW, lr 3e-5 constant, batch 8, 3 epochs (1,746 steps), wd 0.01, fp32, dropout ON in training | | alpha | **0.5** (masks keep a random half) | | seeds | 1004-1023. **Replicas 0-4 (seeds 1004-1008) are Bae's R=5 protocol exactly**; slice `margins[:5, :100]` to compare with published numbers | | masks | M=400 from `sample_subsets(400, 4656, 0.5, 777)` (prefix-stable: masks 0-99 are the original bank's) | | queries | the **full 481-doc validation split** — the battery the 0.468 averages over | | readout | per-query cross-entropy; `margins = -loss` (higher = better); raw `losses` ride along | | built with | bergson @ `0ad2e5b0ce09207d2af0e80383fe2d952dd8e747`, `nproc_per_node: 1`, `--save_models` | ## Two readouts of the same models bergson's `validate` loop at the pinned commit never calls `.eval()` when `train_mode` is on, so the query losses it logs are read **with dropout active**. bergson's published 0.468 re-reads saved models in eval mode. We retrained every (mask, seed) cell with bergson at the pinned commit, saved every model, and read each one ourselves in eval mode (`corrected-lds read-bergson-retrained`) — that is `pool.npz`. bergson's own in-loop readout of the same runs is `pool_dropout_readout.npz`, which also carries the per-replica full-data `baselines`. Dropout at evaluation is most of the seed noise of the in-loop readout (measured on the 100 x 10 bank): | (mean over queries) | `pool.npz` (eval) | `pool_dropout_readout.npz` | |---|---|---| | seed-noise variance, total | 8.78e-05 | 6.99e-04 | | per-replica offset `c(r)` (cancels) | 2.31e-05 (26%) | 6.98e-05 (10%) | | residual (attenuates) | 6.53e-05 | 6.36e-04 | | lambda (median over queries) | **0.37** | **3.86** | The dropout pool reproduces the bank's original pool: for seeds 1004-1008 its cells on masks 0-12 match to ~4e-6, and its 5-seed EK-FAC LDS is 0.3499 vs the original 0.3506. The other cells are fresh dropout draws (bergson's dropout RNG carries across the models of one `validate` call, so only chunks aligned with the original run are bit-identical). ## Results on the 400 x 20 bank (2026-10-25) Every score file on the default eval-readout `pool.npz`, R=20 seeds, M=400 masks (`corrected-lds evaluate`; jackknife SE in brackets; columns are the result JSON's `spearman`, `corrected_rho_inf`, `bridged_rho_inf`, `debiased_noise_corrected_spearman`): | tauq | naive Spearman | corrected rho_inf (SE) | bridged rho_inf | debiased + noise-corrected | |---|---|---|---|---| | `ekfac_bergson` (bergson, 1 model) | 0.4782 | 0.4938 (0.0022) | 0.4782 | 0.4836 | | `ekfac_e10` (ours, 10 models) | 0.4772 | 0.4931 (0.0022) | 0.4776 | 0.4825 | | `ekfac_e5` (ours, 5 models) | 0.4759 | 0.4918 (0.0022) | 0.4763 | 0.4813 | | `source_L3_e1` (bergson's, 1 model) | 0.4723 | 0.4887 (0.0022) | 0.4732 | 0.4776 | | `tracin_e10` (ours, 10 runs) | 0.1267 | 0.1321 (0.0028) | 0.1263 | 0.1282 | | `trak5_e10` (ours, 50 members) | 0.2274 | 0.2380 (0.0025) | 0.2280 | 0.2300 | | `trak1_e50` (ours, 50 members) | 0.2274 | 0.2380 (0.0025) | 0.2281 | 0.2301 | | `trak5_e4` (ours, 20 members) | 0.2079 | 0.2178 (0.0026) | 0.2086 | 0.2103 | | `trak1_e20` (ours, 20 members) | 0.2080 | 0.2179 (0.0026) | 0.2087 | 0.2103 | | `trak5_e2` (ours, 10 members) | 0.1856 | 0.1946 (0.0027) | 0.1863 | 0.1877 | | `trak1_e10` (ours, 10 members) | 0.1856 | 0.1946 (0.0027) | 0.1863 | 0.1877 | With four times the masks the jackknife SE roughly halves; the corrected values sit within about two of the 100 x 10 bank's SEs of the values below (e.g. `ekfac_e10` 0.4931 vs 0.4842, `trak5_e10` 0.2380 vs 0.2397). ## Results on the original 100 x 10 bank (masks 0-99, seeds 1004-1013 of these pools) Mean per-query Spearman across the 100 masks, averaged over the 481 queries (bergson's aggregation), for bergson's EK-FAC, by number of seeds averaged in the readout: | seeds | 1 | 2 | 3 | 4 | **5** | 6 | 7 | 8 | 9 | 10 | |---|---|---|---|---|---|---|---|---|---|---| | eval (`pool.npz`) | .4063 | .4354 | .4445 | .4512 | **.4562** | .4585 | .4609 | .4627 | .4632 | .4637 | | dropout | .2222 | .2789 | .3115 | .3357 | **.3499** | .3585 | .3696 | .3786 | .3856 | .3899 | Every score file, naive LDS vs the corrected estimates (`corrected-lds evaluate`; jackknife SE in brackets): | pool | R | tauq | naive | corrected rho_inf | bridged | debiased | |---|---|---|---|---|---|---| | eval | 10 | `ekfac_bergson` (bergson, 1 model) | 0.4637 | 0.4845 (.0047) | 0.4693 | 0.4763 | | eval | 10 | `ekfac_e10` (ours, 10 models) | 0.4633 | 0.4842 (.0046) | 0.4690 | 0.4758 | | eval | 10 | `ekfac_e5` (ours, 5 models) | 0.4627 | 0.4833 (.0047) | 0.4682 | 0.4752 | | eval | 10 | `source_L3_e1` (bergson's, 1 model) | 0.4622 | 0.4824 (.0047) | 0.4673 | 0.4747 | | eval | 10 | `tracin_e10` (ours, 10 runs) | 0.1194 | 0.1233 (.0056) | 0.1180 | 0.1227 | | eval | 10 | `trak5_e10` (ours, 50 members) | 0.2267 | 0.2397 (.0055) | 0.2298 | 0.2331 | | eval | 10 | `trak1_e50` (ours, 50 members) | 0.2267 | 0.2397 (.0055) | 0.2299 | 0.2331 | | eval | 10 | `trak5_e4` (ours, 20 members) | 0.2102 | 0.2227 (.0054) | 0.2135 | 0.2160 | | eval | 10 | `trak1_e20` (ours, 20 members) | 0.2102 | 0.2228 (.0054) | 0.2136 | 0.2161 | | eval | 10 | `trak5_e2` (ours, 10 members) | 0.1874 | 0.1973 (.0058) | 0.1890 | 0.1927 | | eval | 10 | `trak1_e10` (ours, 10 members) | 0.1872 | 0.1973 (.0058) | 0.1890 | 0.1924 | | eval | 5 | `ekfac_bergson` | **0.4562** | 0.4848 (.0051) | 0.4696 | 0.4764 | | eval | 5 | `ekfac_e10` | 0.4559 | 0.4846 (.0050) | 0.4694 | 0.4761 | | eval | 5 | `ekfac_e5` | 0.4553 | 0.4838 (.0050) | 0.4686 | 0.4755 | | eval | 5 | `source_L3_e1` | 0.4548 | 0.4827 (.0049) | 0.4676 | 0.4750 | | eval | 5 | `tracin_e10` | 0.1165 | 0.1221 (.0060) | 0.1169 | 0.1217 | | eval | 5 | `trak5_e10` | 0.2221 | 0.2388 (.0055) | 0.2290 | 0.2321 | | dropout | 10 | `ekfac_bergson` | 0.3899 | 0.4737 (.0057) | 0.4590 | 0.4633 | | dropout | 10 | `ekfac_e10` | 0.3904 | 0.4743 (.0056) | 0.4596 | 0.4636 | | dropout | 10 | `ekfac_e5` | 0.3896 | 0.4730 (.0056) | 0.4583 | 0.4627 | | dropout | 10 | `source_L3_e1` | 0.3869 | 0.4689 (.0057) | 0.4543 | 0.4591 | | dropout | 10 | `tracin_e10` | 0.1023 | 0.1246 (.0062) | 0.1193 | 0.1217 | | dropout | 10 | `trak5_e10` | 0.2049 | 0.2492 (.0065) | 0.2391 | 0.2442 | | dropout | 5 | `ekfac_bergson` | 0.3499 | 0.4860 (.0071) | 0.4717 | 0.4759 | | dropout | 5 | `ekfac_e10` | 0.3504 | 0.4867 (.0069) | 0.4724 | 0.4759 | | dropout | 5 | `source_L3_e1` | 0.3483 | 0.4813 (.0070) | 0.4672 | 0.4730 | | dropout | 5 | `tracin_e10` | 0.0909 | 0.1245 (.0078) | 0.1193 | 0.1226 | | dropout | 5 | `trak5_e10` | 0.1834 | 0.2574 (.0070) | 0.2473 | 0.2513 | The two readouts disagree by 0.07-0.11 in naive LDS and agree to ~0.01 once corrected: the dropout readout is the same signal under ten times the noise, and the correction recovers it. **Published references.** bergson's example README: EK-FAC IF **0.468 ± 0.015** (eval readout, five seeds), and "kronfluence reports 0.44". Bae et al. 2024 print no WikiText-2 numbers; their Fig. 6 bars (alpha 0.5, R=5; read off the figure, ±0.005) are IF 0.43 (one model) / 0.445 (ten models), SOURCE 0.44 / 0.47, TracIn 0.10, RepSim 0.01. Our EK-FAC (kronfluence) matches bergson's to ≤0.001 on every pool. Its corrected LDS on the eval pool is flat in the number of runs averaged: 0.4687 (1 run), 0.4833 (5), 0.4842 (10), so ten of our runs land level with bergson's single model (0.4845). **SOURCE (one model) scores 0.4548 on Bae's five seeds (0.4622 at R=10)**, against bergson's 0.475 ± 0.015 and Bae's ≈0.44. On the same model (seed 30000) it beats our EK-FAC by +0.0135 (paired SE 0.0018), Bae's single-model ordering (SOURCE 0.44 > IF 0.43), and it beats each of our ten single EK-FAC runs (0.447-0.453); one SOURCE model matches our ten-run EK-FAC (0.4633). It is level with bergson's single-model EK-FAC (0.4637, difference -0.0015 ± 0.0023), which scores above every one of our single EK-FAC runs. **Our TracIn scores 0.1165 on Bae's five seeds (0.1194 at R=10), against Bae's ≈0.10.** Like theirs it does not improve with more runs: corrected 0.1222 for one run, 0.1233 for ten. The TracIn this bank shipped before scored 0.055: it traced all 124M parameters, where the tied `wte`/`lm_head` carries ~91% of the gradient norm, and one 1,024-dim projection per run lost half of what remained. **Our TRAK scores 0.2221 on Bae's five seeds (0.2267 at R=10; corrected 0.2397) with 50 members.** Bae et al. omit TRAK on WikiText ("no public LM implementation"), so there is no published reference. Unlike EK-FAC and TracIn here it keeps rising with members -- corrected 0.083 for one member, 0.169 for one unit of five, 0.223 at 20, 0.240 at 50 -- and stays well below EK-FAC. `trak5` and `trak1` agree to 1e-4 at equal members. ## Two properties that make this bank unlike the others here **1. `seed_mode = "shared-per-replica"`.** bergson uses ONE seed per replica, shared across that replica's masks, where our other banks seed every *cell*. The readout decomposes as `y[r,j,q] = mu(S_j) + c(r) + eps(r,j)`, and the per-replica offset `c(r)` shifts a whole column, so it **cancels** in a correlation across masks and does not attenuate rho. `evaluate_pool` reads the stamp and uses `estimator.seed_noise_variance_centred` for such a pool (corrected-lds `main` since 2026-10-06). **2. Masks are applied by zeroing loss weights, not by retraining on the subset.** Batching and data order are identical across masks — a *different counterfactual* from our CIFAR/QNLI/ImageNet banks. A companion measurement, not another entry in that series. ## Loading Listed in corrected-lds's `banks-manifest.json`, which pins every file's sha256: ```bash corrected-lds fetch-banks --bank wikitext2_gpt2_bae corrected-lds evaluate banks/wikitext2_gpt2_bae --tauq ekfac_e10 # eval readout corrected-lds evaluate banks/wikitext2_gpt2_bae/pool_dropout_readout.npz --tauq ekfac_e10 ``` ```python from corrected_lds import load_pool, load_tauq from corrected_lds.estimator import seed_noise_variance_centred, seed_offset_variance pool = load_pool("wikitext2_gpt2_bae/pool.npz") # eval readout, shared-per-replica tau = load_tauq("wikitext2_gpt2_bae/tauq_ekfac_bergson.npz") resid = seed_noise_variance_centred(pool.margins) # the part that attenuates offset = seed_offset_variance(pool.margins) # the part that does not ``` Identity is enforced on load: the stamped side digests must equal the `wikitext2` pins in `banks/presets.py`, which hash the int64 `input_ids` of the pinned dataset revision. The same revision is passed to bergson via `data.data_kwargs`, so both halves provably read identical bytes. Both pools carry the same masks and examples, so every tauq and pool_tda here pairs with either. ## Build record How to rebuild the bank, with each file's verification status, is in `docs/REPRODUCE.md` in the corrected-lds repository.