Initial upload: MAGIC scores for GPT-2 / wikitext-2 / test[3:4] query

README.md

@@ -0,0 +1,113 @@
+---
+license: mit
+language:
+- en
+tags:
+- influence-functions
+- attribution
+- magic
+- bergson
+- gpt-2
+- wikitext
+---
+
+# bergson MAGIC scores — GPT-2 / wikitext-2 / random validation
+
+Per-document MAGIC attribution scores for a GPT-2 model fine-tuned on
+`Salesforce/wikitext` `wikitext-2-raw-v1` `train`, attributing the
+training-loss gradient with respect to the model's output on
+`test[3:4]` (a single test example, taken as the query).
+
+Scores were produced by the [`bergson`](https://github.com/EleutherAI/bergson)
+attribution toolkit's MAGIC subcommand, which back-propagates through
+the entire training trajectory to compute, for each training doc `d`,
+`∂(L_query)/∂(w_d)` — the first-order Taylor expansion of how much
+removing doc `d` from training would change the query loss.
+
+## Files
+
+| file | description |
+|---|---|
+| `scores.pt` | torch.float32 tensor, shape `(36718,)`. Indexed by original wikitext-2 train row position (filtered <2-token rows have score 0). |
+| `summary.csv` | Final Spearman/Pearson correlation between MAGIC scores and 100-subset leave-k-out training-loss diffs at validate time. |
+| `validation.csv` | Per-subset (`subset, diff, score_sum`) records from the leave-k-out validation. |
+
+## Loading
+
+```python
+import torch
+scores = torch.load("scores.pt", map_location="cpu", weights_only=True)
+# scores.shape == (36718,), one entry per original wikitext-2 train row.
+```
+
+## Validation result
+
+Random-strategy 100-subset leave-k-out validation, batch_size=256,
+num_epochs=2, polynomial LR schedule:
+
+| metric | value | p |
+|---|---|---|
+| Spearman ρ | **+0.9731** | 2.71e-64 |
+| Pearson r | **+0.9724** | 1.00e-63 |
+| baseline_loss | 3.6840 |  |
+
+(Sorted-strategy validation against the same scores reaches ρ = +0.9926
+on the same 100 subsets.)
+
+## YAML used to generate these scores
+
+```yaml
+- magic:
+    run_path: runs/compare/q3_random
+    model: gpt2
+    overwrite: true
+
+    data:
+      dataset: Salesforce/wikitext
+      subset: wikitext-2-raw-v1
+      split: "train"
+      chunk_length: 512
+
+    query:
+      dataset: Salesforce/wikitext
+      subset: wikitext-2-raw-v1
+      split: "test[3:4]"
+      chunk_length: 0
+
+    distributed:
+      nproc_per_node: 4
+      nnode: 4
+
+    batch_size: 256
+    num_epochs: 2
+    lr_schedule:
+      lr_scheduler_type: polynomial
+      lr: 0.0008
+      lr_start: 1e-6
+      lr_end: 0.00008
+      warmup_steps: 0.25
+
+    subset_strategy: random
+    wandb_project: magic
+```
+
+Saved as `examples/magic/compare/q3_magic.yaml` in the bergson repo
+(the consolidated pipeline that produces both random and sorted
+validates from a single MAGIC backward pass).
+
+Run with:
+
+```bash
+bergson pipeline examples/magic/compare/q3_magic.yaml
+```
+
+## Citation
+
+```
+@misc{bergson,
+  title  = {bergson: gradient-based attribution for language models},
+  author = {EleutherAI},
+  year   = {2026},
+  url    = {https://github.com/EleutherAI/bergson},
+}
+```

scores.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b0cd23a1ed53ddc38db3dd474c7be9c9db551df75784cfa976567e61a0552d54
+size 148386

summary.csv

@@ -0,0 +1,2 @@
+spearman_corr,spearman_p,pearson_corr,pearson_p,N,baseline_loss
+0.973105310531053,2.7091898753024367e-64,0.9723677280660237,1.0022088934671183e-63,100,3.6840500831604004

validation.csv

@@ -0,0 +1,101 @@
+subset,diff,score_sum
+0,-0.0018310546875,-0.029189864173531532
+1,-0.0013134479522705078,-0.022047167643904686
+2,-0.0009467601776123047,-0.01643146015703678
+3,0.0018296241760253906,0.033269889652729034
+4,0.0031800270080566406,0.04270397126674652
+5,-0.0015027523040771484,-0.01642679050564766
+6,0.0008563995361328125,0.00962444394826889
+7,0.0005595684051513672,0.016617674380540848
+8,-0.0004363059997558594,-0.006987039931118488
+9,0.001634836196899414,0.02674231119453907
+10,-3.6716461181640625e-05,-0.0027552940882742405
+11,0.002792835235595703,0.0499853640794754
+12,-3.0994415283203125e-06,0.003259955905377865
+13,0.0001537799835205078,0.005294406320899725
+14,0.0016977787017822266,0.024854004383087158
+15,0.0017843246459960938,0.03346139192581177
+16,0.0012364387512207031,0.021332142874598503
+17,-0.002604246139526367,-0.03763214871287346
+18,-0.001325368881225586,-0.022946178913116455
+19,0.0031244754791259766,0.048977550119161606
+20,0.001478433609008789,0.02487548254430294
+21,-0.0016148090362548828,-0.030084803700447083
+22,-0.0020041465759277344,-0.02288348600268364
+23,-0.0005414485931396484,-0.006325895432382822
+24,-0.0022764205932617188,-0.02015402913093567
+25,-0.00021648406982421875,-0.006860801950097084
+26,0.0009942054748535156,0.011761580593883991
+27,-0.0025680065155029297,-0.04330901801586151
+28,-0.0010652542114257812,-0.020170079544186592
+29,0.0003962516784667969,0.006441427394747734
+30,0.0008099079132080078,0.01868048869073391
+31,-0.0024933815002441406,-0.037735529243946075
+32,-0.00014138221740722656,-0.0009743906557559967
+33,-0.002919435501098633,-0.04292494058609009
+34,0.00034928321838378906,0.005179773084819317
+35,-0.0002903938293457031,0.007907246239483356
+36,0.0010368824005126953,0.019828442484140396
+37,-0.0028319358825683594,-0.025016576051712036
+38,0.0005180835723876953,0.004872254561632872
+39,0.0026693344116210938,0.035355955362319946
+40,-0.0001232624053955078,-0.011232763528823853
+41,0.0009353160858154297,0.025380903854966164
+42,0.002220630645751953,0.032907065004110336
+43,0.0010669231414794922,0.01819712668657303
+44,-0.0002465248107910156,0.0034509990364313126
+45,0.0021257400512695312,0.04210555553436279
+46,-0.002271413803100586,-0.04287688434123993
+47,-0.0014812946319580078,-0.016457801684737206
+48,0.0010526180267333984,0.01314145140349865
+49,0.0016613006591796875,0.030897097662091255
+50,0.0010073184967041016,0.015840895473957062
+51,0.00032019615173339844,0.006934883072972298
+52,-0.0007016658782958984,-0.012217387557029724
+53,-0.0004940032958984375,-0.007514691445976496
+54,0.0003879070281982422,0.010312681086361408
+55,0.001552581787109375,0.02759663015604019
+56,0.0004470348358154297,0.004183738026767969
+57,0.0021224021911621094,0.03304430469870567
+58,-0.0004887580871582031,0.005980587564408779
+59,0.0019378662109375,0.03738723322749138
+60,-0.0011758804321289062,-0.018757693469524384
+61,-0.00015592575073242188,0.004054620862007141
+62,0.0011279582977294922,0.01150202751159668
+63,0.0019664764404296875,0.03209429234266281
+64,0.00013780593872070312,0.010195817798376083
+65,0.001668691635131836,0.02927013300359249
+66,0.0007617473602294922,0.004966499749571085
+67,0.0011334419250488281,0.023897934705018997
+68,-0.0010595321655273438,-0.0056649986654520035
+69,-0.002534627914428711,-0.033193040639162064
+70,0.0011744499206542969,0.02143128588795662
+71,-0.0005335807800292969,-0.0032467744313180447
+72,0.0008752346038818359,0.015030361711978912
+73,-0.0014462471008300781,-0.03063521906733513
+74,-0.00019431114196777344,-0.000977933406829834
+75,-0.0014472007751464844,-0.005219193175435066
+76,-0.002418994903564453,-0.02468314953148365
+77,-0.0003635883331298828,-0.002361757680773735
+78,-0.0039768218994140625,-0.057306695729494095
+79,-0.0016314983367919922,-0.03249344602227211
+80,0.00024127960205078125,0.005457842722535133
+81,0.0016019344329833984,0.022662054747343063
+82,-0.0007393360137939453,-0.012122330255806446
+83,0.0028350353240966797,0.049485936760902405
+84,0.0005719661712646484,-0.0009666034020483494
+85,-0.002422332763671875,-0.03476095199584961
+86,-0.0012798309326171875,-0.014440640807151794
+87,0.00026607513427734375,0.0011126925237476826
+88,0.00023508071899414062,0.004115729592740536
+89,0.0020787715911865234,0.028063565492630005
+90,0.001421213150024414,0.025426872074604034
+91,0.0023102760314941406,0.0365045890212059
+92,-0.0005500316619873047,-0.013258624821901321
+93,0.0008111000061035156,0.016341816633939743
+94,-0.00020933151245117188,-0.002287695650011301
+95,0.0021834373474121094,0.03424829989671707
+96,-1.3589859008789062e-05,0.0005636848509311676
+97,0.0012981891632080078,0.012789108790457249
+98,0.0011246204376220703,0.021215496584773064
+99,0.0002624988555908203,0.005299761425703764