File size: 16,514 Bytes
708a9b9
8b4b1f4
 
a95a860
8b4b1f4
708a9b9
 
 
 
 
 
 
 
 
686d467
708a9b9
686d467
 
 
 
 
 
708a9b9
686d467
f9b6c4b
686d467
 
 
708a9b9
a95a860
b3ad680
686d467
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a95a860
686d467
 
b3ad680
 
686d467
 
a95a860
 
 
708a9b9
686d467
 
 
 
 
 
 
 
 
 
 
 
a95a860
708a9b9
a95a860
 
 
 
 
 
 
adb315b
a95a860
 
 
 
 
708a9b9
a95a860
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
708a9b9
a95a860
adb315b
a95a860
 
 
708a9b9
 
 
 
adb315b
 
 
 
 
 
 
 
 
a95a860
708a9b9
a95a860
 
 
8b4b1f4
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
---
datasets:
- hematteo/wu-crosscoder-snapshots
license: mit
pipeline_tag: other
tags:
- crosscoder
- sparse-autoencoder
- mech-interp
- parameter-trajectory
- pythia
- olmo
---

# Parameter-trajectory crosscoders for the LM readout

Trained dictionaries from **Learning to Read Out: Unembedding Dynamics in Language
Model Pretraining** (NeurIPS 2026). A parameter-trajectory crosscoder fits one
sparse dictionary to every token's readout (unembedding, $W_U$) row at every
checkpoint, using only checkpoint weights. Each feature keeps one index across
checkpoints and has its own decoder direction at each, so you can follow how its
strength and the tokens it promotes change during pretraining.

[Paper (arXiv:2610.21025)](https://arxiv.org/abs/2610.21025) 路
[Project page](https://matteohe.com/learning-to-read-out/) 路
[Feature atlas](https://huggingface.co/spaces/hematteo/readout-feature-atlas) 路
[Code](https://github.com/hematteo/learning-to-read-out) 路
[Collection](https://huggingface.co/collections/hematteo/learning-to-read-out-neurips-2026-6ad25219e26a69daa47880e0)

## Quick start

Follow one feature of the Pythia-160M dictionary across training (downloads
1.6 GB for the dictionary and 155 MB for one readout snapshot):

```python
import json
import torch
from huggingface_hub import hf_hub_download
from safetensors import safe_open
from transformers import AutoTokenizer

repo = "hematteo/parameter-trajectory-crosscoders"
name = "pythia-160m/W_U/cross-snapshot-32/d8192/seed0"
weights = hf_hub_download(repo, f"{name}.safetensors")
steps = json.load(open(hf_hub_download(repo, f"{name}.config.json")))["steps"]  # the 32 checkpoint steps

feature = 6046
with safe_open(weights, framework="pt") as f:
    D = f.get_slice("W_D")[:, feature]  # (32, d_model): this feature's decoder row at each checkpoint
rho = D.norm(dim=-1) / D.norm(dim=-1).max()  # decoder norm relative to its peak
print({s: round(rho[i].item(), 2) for i, s in enumerate(steps) if s in (0, 512, 2000, 8000, 143000)})
# {0: 0.05, 512: 0.07, 2000: 0.44, 8000: 0.78, 143000: 0.93}

snap = hf_hub_download("hematteo/wu-crosscoder-snapshots",
                       "pythia-160m/EleutherAI_pythia-160m_step143000_wu.pt", repo_type="dataset")
W_U = torch.load(snap, weights_only=True)  # (vocab, d_model) readout at step 143000
tok = AutoTokenizer.from_pretrained("EleutherAI/pythia-160m")
top = (W_U @ D[steps.index(143000)]).topk(10).indices
print([tok.decode([t]) for t in top])
# [' tumour', ' realise', ' characterised', ' realised', ' colour', ' recognise', ' tumours', ' whilst', ' recognised', ' favour']
```

Feature 6046 promotes British spellings at the end of training, and its decoder
norm grows from 7% of its peak at step 512 to 93% at the final checkpoint. The
[feature atlas](https://huggingface.co/spaces/hematteo/readout-feature-atlas)
has a page like this for every feature of the Pythia-160M, 1B and 6.9B
dictionaries.

To download whole dictionaries:

```bash
hf download hematteo/parameter-trajectory-crosscoders --include "pythia-1b/**" --local-dir ./parameter-trajectory-crosscoders
hf download hematteo/parameter-trajectory-crosscoders --local-dir ./parameter-trajectory-crosscoders  # everything, about 180 GB
```

## Files

Each artifact is `<name>.safetensors` + `<name>.config.json` (training
hyperparameters and recomputed quality metrics) + `<name>.md` (card).
`index.json` is the machine-readable inventory of everything below.

| Tensor | Shape | Meaning |
|---|---|---|
| `W_D` | (checkpoints, d_sae, d_model) | Decoder: one direction per feature and checkpoint |
| `b_D` | (checkpoints, d_model) | Decoder bias per checkpoint |
| `W_E` | (checkpoints, d_model, d_sae) | Encoder (unrelated to the input embedding $W_E$ below) |
| `b_E` | (checkpoints, d_sae) | Encoder bias |
| `activation_function.log_jumprelu_threshold` | (d_sae,) | JumpReLU threshold (log scale) |

The checkpoint steps are listed under `steps` in each `.config.json`. Inputs were
centered and scaled per checkpoint (`preprocess_mode` in the config); the code
release reproduces that preprocessing.

## What you probably want

| What | Path |
|---|---|
| Headline 5-seed Pythia-160M $W_U$ crosscoder | `pythia-160m/W_U/cross-snapshot-32/d8192/seed{0..4}.safetensors` |
| High-resolution 160M instrument (atlas) | `pythia-160m/W_U/cross-snapshot-32/d24576/seed0.safetensors` |
| Cross-scale (Pythia-1B) | `pythia-1b/W_U/cross-snapshot-32/d24576/seed0.safetensors` |
| Large-scale, selected sparse run | `pythia-6.9b/W_U/cross-snapshot-32/d32768/seed0-sparse.safetensors` |
| Cross-family (OLMo-2-7B) | `olmo-2-7b/W_U/cross-snapshot-32/d32768/seed0.safetensors` |
| Input embedding $W_E$ (released; the paper's analyses use only $W_U$) | `pythia-160m/W_E/cross-snapshot-32/...` |
| Activation-rate aggregates (lifecycle figures) | `derived/aggregates/`, `derived/rates/` |
| Attribution-patching artifacts | `attribution/pythia-160m/` |
| Held-out eval token corpus | `evaluation/eval-corpus/eval_tokens.pt` |

## Full inventory

| Path | Model | Matrix | Kind | d_sae | Seed | Quality |
|---|---|---|---|---|---|---|
| `olmo-2-7b/W_U/cross-snapshot-32/d32768/seed0.safetensors` | allenai/OLMo-2-1124-7B | W_U | cross-snapshot-32 | 32768 | 0 | EV 0.853 / L0 557 |
| `pythia-160m/W_E/cross-snapshot-32/d24576/seed0.safetensors` | EleutherAI/pythia-160m | W_E | cross-snapshot-32 | 24576 | 0 | EV 0.831 / L0 118 |
| `pythia-160m/W_E/cross-snapshot-32/d8192/seed0.safetensors` | EleutherAI/pythia-160m | W_E | cross-snapshot-32 | 8192 | 0 | EV 0.581 / L0 82 |
| `pythia-160m/W_E/cross-snapshot-32/d8192/seed1.safetensors` | EleutherAI/pythia-160m | W_E | cross-snapshot-32 | 8192 | 1 | EV 0.580 / L0 82 |
| `pythia-160m/W_E/cross-snapshot-32/d8192/seed2.safetensors` | EleutherAI/pythia-160m | W_E | cross-snapshot-32 | 8192 | 2 | EV 0.582 / L0 82 |
| `pythia-160m/W_E/cross-snapshot-32/d8192/seed3.safetensors` | EleutherAI/pythia-160m | W_E | cross-snapshot-32 | 8192 | 3 | EV 0.581 / L0 82 |
| `pythia-160m/W_E/cross-snapshot-32/d8192/seed4.safetensors` | EleutherAI/pythia-160m | W_E | cross-snapshot-32 | 8192 | 4 | EV 0.583 / L0 83 |
| `pythia-160m/W_U/architecture-comparison/d8192/batchtopk/seed0.safetensors` | EleutherAI/pythia-160m | W_U | architecture-comparison/d8192 | 8192 | 0 | EV 0.725 / L0 203 |
| `pythia-160m/W_U/architecture-comparison/d8192/gated/seed0.safetensors` | EleutherAI/pythia-160m | W_U | architecture-comparison/d8192 | 8192 | 0 | EV 0.214 / L0 12 |
| `pythia-160m/W_U/architecture-comparison/d8192/gated-retuned/seed0.safetensors` | EleutherAI/pythia-160m | W_U | architecture-comparison/d8192 | 8192 | 0 | EV 0.827 / L0 654 |
| `pythia-160m/W_U/cross-snapshot-16/d8192/seed0.safetensors` | EleutherAI/pythia-160m | W_U | cross-snapshot-16 | 8192 | 0 | EV 0.773 / L0 216 |
| `pythia-160m/W_U/cross-snapshot-32/d16384/seed0.safetensors` | EleutherAI/pythia-160m | W_U | cross-snapshot-32 | 16384 | 0 | EV 0.780 / L0 103 |
| `pythia-160m/W_U/cross-snapshot-32/d24576/seed0.safetensors` | EleutherAI/pythia-160m | W_U | cross-snapshot-32 | 24576 | 0 | EV 0.920 / L0 286 |
| `pythia-160m/W_U/cross-snapshot-32/d24576/seed1.safetensors` | EleutherAI/pythia-160m | W_U | cross-snapshot-32 | 24576 | 1 | EV 0.920 / L0 286 |
| `pythia-160m/W_U/cross-snapshot-32/d24576/seed2.safetensors` | EleutherAI/pythia-160m | W_U | cross-snapshot-32 | 24576 | 2 | EV 0.920 / L0 286 |
| `pythia-160m/W_U/cross-snapshot-32/d8192/seed0.safetensors` | EleutherAI/pythia-160m | W_U | cross-snapshot-32 | 8192 | 0 | EV 0.776 / L0 203 |
| `pythia-160m/W_U/cross-snapshot-32/d8192/seed1.safetensors` | EleutherAI/pythia-160m | W_U | cross-snapshot-32 | 8192 | 1 | EV 0.776 / L0 203 |
| `pythia-160m/W_U/cross-snapshot-32/d8192/seed2.safetensors` | EleutherAI/pythia-160m | W_U | cross-snapshot-32 | 8192 | 2 | EV 0.776 / L0 203 |
| `pythia-160m/W_U/cross-snapshot-32/d8192/seed3.safetensors` | EleutherAI/pythia-160m | W_U | cross-snapshot-32 | 8192 | 3 | EV 0.776 / L0 203 |
| `pythia-160m/W_U/cross-snapshot-32/d8192/seed4.safetensors` | EleutherAI/pythia-160m | W_U | cross-snapshot-32 | 8192 | 4 | EV 0.777 / L0 203 |
| `pythia-160m/W_U/final-snapshot-saes/d16384.safetensors` | EleutherAI/pythia-160m | W_U | final-snapshot-saes | 16384 | 0 | EV 0.870 / L0 1913 |
| `pythia-160m/W_U/final-snapshot-saes/d32768.safetensors` | EleutherAI/pythia-160m | W_U | final-snapshot-saes | 32768 | 0 | EV 0.926 / L0 3410 |
| `pythia-160m/W_U/final-snapshot-saes/d6144.safetensors` | EleutherAI/pythia-160m | W_U | final-snapshot-saes | 6144 | 0 | EV 0.765 / L0 862 |
| `pythia-160m/W_U/final-snapshot-saes/d65536.safetensors` | EleutherAI/pythia-160m | W_U | final-snapshot-saes | 65536 | 0 | EV 0.964 / L0 5943 |
| `pythia-160m/W_U/final-snapshot-saes/d8192.safetensors` | EleutherAI/pythia-160m | W_U | final-snapshot-saes | 8192 | 0 | EV 0.799 / L0 1084 |
| `pythia-160m/W_U/lambda-sweep/d8192/lam0p40_seed0.safetensors` | EleutherAI/pythia-160m | W_U | lambda-sweep | 8192 | 0 | EV 0.748 / L0 160 |
| `pythia-160m/W_U/lambda-sweep/d8192/lam1p00_seed0.safetensors` | EleutherAI/pythia-160m | W_U | lambda-sweep | 8192 | 0 | EV 0.632 / L0 58 |
| `pythia-160m/W_U/lambda-sweep/d8192/lam1p20_seed0.safetensors` | EleutherAI/pythia-160m | W_U | lambda-sweep | 8192 | 0 | EV 0.603 / L0 45 |
| `pythia-160m/W_U/lambda-sweep/d8192/lam1p35_seed0.safetensors` | EleutherAI/pythia-160m | W_U | lambda-sweep | 8192 | 0 | EV 0.582 / L0 38 |
| `pythia-160m/W_U/lambda-sweep/d8192/lam1p35_seed1.safetensors` | EleutherAI/pythia-160m | W_U | lambda-sweep | 8192 | 1 | EV 0.582 / L0 38 |
| `pythia-160m/W_U/lambda-sweep/d8192/lam1p35_seed2.safetensors` | EleutherAI/pythia-160m | W_U | lambda-sweep | 8192 | 2 | EV 0.582 / L0 38 |
| `pythia-160m/W_U/lambda-sweep/d8192/lam1p80_seed0.safetensors` | EleutherAI/pythia-160m | W_U | lambda-sweep | 8192 | 0 | EV 0.528 / L0 23 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step0.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.729 / L0 1150 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step1.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.729 / L0 1150 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step1000.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.786 / L0 997 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step102000.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.800 / L0 983 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step116000.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.812 / L0 958 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step128.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.729 / L0 1150 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step130000.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.820 / L0 940 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step14000.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.791 / L0 996 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step143000.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.824 / L0 924 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step16.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.729 / L0 1150 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step2.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.729 / L0 1150 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step2000.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.799 / L0 969 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step21000.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.791 / L0 998 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step256.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.732 / L0 1142 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step27000.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.791 / L0 999 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step3000.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.799 / L0 972 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step32.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.729 / L0 1150 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step34000.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.791 / L0 1000 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step4.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.729 / L0 1150 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step4000.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.798 / L0 977 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step47000.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.791 / L0 1000 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step5000.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.796 / L0 982 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step512.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.754 / L0 1087 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step6000.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.795 / L0 985 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step61000.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.790 / L0 1002 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step64.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.729 / L0 1150 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step7000.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.794 / L0 988 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step75000.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.790 / L0 1004 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step8.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.729 / L0 1150 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step8000.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.793 / L0 990 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step89000.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.792 / L0 1002 |
| `pythia-160m/W_U/per-snapshot-saes/d8192/step9000.safetensors` | EleutherAI/pythia-160m | W_U | per-snapshot-saes | 8192 | 0 | EV 0.792 / L0 992 |
| `pythia-1b/W_U/cross-snapshot-32/d16384/seed0.safetensors` | EleutherAI/pythia-1b | W_U | cross-snapshot-32 | 16384 | 0 | EV 0.781 / L0 499 |
| `pythia-1b/W_U/cross-snapshot-32/d24576/seed0.safetensors` | EleutherAI/pythia-1b | W_U | cross-snapshot-32 | 24576 | 0 | EV 0.861 / L0 517 |
| `pythia-1b/W_U/cross-snapshot-32/d8192/seed0.safetensors` | EleutherAI/pythia-1b | W_U | cross-snapshot-32 | 8192 | 0 | EV 0.628 / L0 374 |
| `pythia-1b/W_U/cross-snapshot-32-matched-window/d24576/seed0.safetensors` | EleutherAI/pythia-1b | W_U | cross-snapshot-32-matched-window | 24576 | 0 | EV 0.884 / L0 264 |
| `pythia-6.9b/W_U/cross-snapshot-32/d32768/seed0-sparse.safetensors` | EleutherAI/pythia-6.9b | W_U | cross-snapshot-32 | 32768 | 0 | EV 0.808 / L0 742 |
| `pythia-6.9b/W_U/cross-snapshot-32/d32768/seed0.safetensors` | EleutherAI/pythia-6.9b | W_U | cross-snapshot-32 | 32768 | 0 | EV 0.833 / L0 1957 |

Quality metrics are recomputed from the released weights on the released
snapshot schedule (see the code repo's `reproduce/validation/recompute_quality.py`).
The `gated` architecture-comparison run intentionally documents
位-transfer failure (default 位=0.3 moved across architectures); see
`gated-retuned` (位=0.05) for the tuned comparison point.

## Citation

```bibtex
@inproceedings{he2026learningtoreadout,
  title     = {Learning to Read Out: Unembedding Dynamics in Language Model Pretraining},
  author    = {He, Matteo and Shen, William F. and Iacob, Alex and Jovanovic, Andrej
               and Qiu, Xinchi and Lane, Nicholas D.},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026},
  eprint    = {2610.21025},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
}
```

MIT. W_U/W_E source tensors derive from public Apache-2.0 checkpoints
(EleutherAI Pythia, AllenAI OLMo-2). The eval corpus derives from Wikipedia
(CC-BY-SA 4.0).