Ionel2023/read-locate-rewrite-replication

Read · Locate · Rewrite: replication results

These are the full results of a replication of two interpretability papers, made for a two-hour MIT workshop:

• SAE probing. Kantamneni, Engels, Rajamanoharan, Tegmark & Nanda, Are Sparse Autoencoders Useful? A Case Study in Sparse Probing, ICML 2025.
• ROME. Meng, Bau, Andonian & Belinkov, Locating and Editing Factual Associations in GPT, NeurIPS 2022.

Every claim in both papers was tested: 86 claims, plus 23 checks of our own. The one exception is ROME's human-rater study, which cannot be replicated without raters.

The verdicts are:

• SAE probing: 39 of 51 claims hold, 9 partly hold, and 3 do not.
• ROME: 26 of 35 claims hold, 8 partly hold, and 1 is not replicable.

CLAIMS_TABLE.md is the one-page summary. The per-claim ledgers are sae/CLAIMS.md and rome/CLAIMS.md. Each gives the test, the numbers and the result files, and the paths in them are relative to this repository.

The workshop's participant repo is github.com/nelionel/read-locate-rewrite.

Download

On a workshop pod, or anywhere:

# everything (about 10 GB)
huggingface-cli download Ionel2023/read-locate-rewrite-replication --repo-type dataset --local-dir /workspace/replication
# only the ledgers and results (about 2.7 GB)
huggingface-cli download Ionel2023/read-locate-rewrite-replication --repo-type dataset --local-dir /workspace/replication \
  --include "*.md" "sae/results/*" "rome/results/*" "rome/figures/*"

In the participant repo, python tools/get_replication_data.py does the same.

What is here

Path	Contents	Size
CLAIMS_TABLE.md, REPLICATION.md	The ranked claim table, and the replication's overview and compute log	small
sae/CLAIMS.md, rome/CLAIMS.md, rome/REPORT.md	Per-claim ledgers: test, numbers and result files	small
sae/results/	Probe results for every regime: normal, scarcity, imbalance, label noise, covariate shift, multi-token, Llama checks, and our own checks DX‑1 to DX‑14	2.4 GB
sae/activations/gemma-2-9b/<task>_blocks.20.hook_resid_post.pt	Gemma‑2‑9B residual stream after layer 20, at the last prompt token, for the 113 binary tasks, plus out-of-distribution sets. float32 tensors [n_prompts, 3584], in the order of the authors' task files	4.5 GB
sae/encodings/<sae>/<task>.pt	The same activations encoded by four Gemma Scope SAEs (16k, 131k at two L0s, 1M)	2.9 GB
sae/labels/<task>.npz	y (0/1), and the authors' seed-42 split train_idx / test_idx, built as in sae/scripts/r1_probes.py. index.csv lists n per task	small
rome/results/	Causal traces (1,000 known facts, three corruption kinds, windows, severing), CounterFact evaluations (GPT‑2 XL 3,036 edits, GPT‑2 M/L, GPT‑J 2,000, zsRE), design-choice ablations, and our own checks DXR‑1 to DXR‑9	251 MB
rome/figures/	Figures made from the results	4 MB
sae/scripts/, rome/code/, coordinator/, common/	The code that produced everything	6 MB
sae/logs/, rome/logs/	Run logs	48 MB

Not included:

• the Gemma Scope SAE weights, which are on Hugging Face at google/gemma-scope-9b-pt-res;
• the authors' task files with the prompt texts, which you can get from the authors' data release (see JoshEngels/SAE-Probes);
• the model weights.

Load one task

import numpy as np, torch
from huggingface_hub import hf_hub_download
R, T = "Ionel2023/read-locate-rewrite-replication", "92_glue_sst2"
X = torch.load(hf_hub_download(R, f"sae/activations/gemma-2-9b/{T}_blocks.20.hook_resid_post.pt", repo_type="dataset"))[:5000].numpy()
L = np.load(hf_hub_download(R, f"sae/labels/{T}.npz", repo_type="dataset"))
y, tr, te = L["y"], L["train_idx"], L["test_idx"]

Logistic regression on X[tr] gives a test AUC of 0.952 on this task.

Licences and sources

• Our results, ledgers and code: CC BY 4.0 for the results and text. The code is also offered under the MIT licence.
• Activations and encodings: these are outputs of Gemma‑2‑9B and of the Gemma Scope SAEs (CC BY 4.0). Using them is subject to the Gemma Terms of Use.
• Labels: derived from the authors' task files, which compile public datasets. See their paper for each source.
• ROME: experiments use the authors' code (kmeng01/rome, MIT), CounterFact, and the known-facts set from rome.baulab.info.

If you use these results, please cite the two papers above.