issues
search
jbloomAus
/
SAELens
Training Sparse Autoencoders on Language Models
https://jbloomaus.github.io/SAELens/
MIT License
481
stars
127
forks
source link
issues
Newest
Newest
Most commented
Recently updated
Oldest
Least commented
Least recently updated
[Bug Report] Unusual L0 Behavior with Gemma-2b
#382
muyo8692
opened
1 day ago
0
[Bug Report] fold_activation_norm_scaling_factor should be performed before saving all checkpoints, not only the final checkpoint
#381
chanind
opened
2 days ago
0
fix: remove typeguard dependency
#380
chanind
closed
3 days ago
1
chore: adding test that all config params pass to sae
#379
chanind
closed
1 week ago
1
Added new SAEBench gemma 2 2b SAEs
#378
curt-tigges
closed
1 week ago
1
fix: hotfix scale decoder norm is not passed to training sae
#377
chanind
closed
1 week ago
1
Fix parameter naming: Separate l0_lambda from l1_coefficient
#376
muyo8692
closed
1 day ago
5
fix: add neuronpedia ids for llamascope
#375
hijohnnylin
closed
1 week ago
1
chore: fixing whitespace so docs render as list not paragraph
#374
chanind
closed
1 week ago
1
fix: fixing jumprelu encode and save/load
#373
chanind
closed
1 week ago
2
chore: add codecov.yaml and exclude legacy files
#372
chanind
closed
1 week ago
2
fixed llamascope sae names and loader
#371
curt-tigges
closed
1 week ago
1
feat: Topk SAE training
#370
chanind
closed
3 days ago
1
Add Llama Scope SAEs & improvements to evaluating ce scores.
#369
Hzfinfdu
closed
1 week ago
3
[Proposal] Move to Ruff for linting / formatting
#368
chanind
opened
1 week ago
3
Faster cleanup of sharded datasets
#367
tom-pollak
closed
1 week ago
2
chore: updating training docs with tips / jumprelu
#366
chanind
closed
1 week ago
2
fix: only scale sparsity by dec norm if specified in the config
#365
chanind
closed
1 week ago
2
I have checked that there is no similar issue in the repo (required)
#364
chanind
closed
2 weeks ago
1
[Bug Report] scale_sparsity_penalty_by_decoder_norm option is ignored
#363
chanind
closed
1 week ago
0
chore: adding 'Load this SAE' popup to docs table
#362
chanind
closed
2 weeks ago
2
fix: load the same config from_pretrained and get_sae_config
#361
chanind
closed
2 weeks ago
3
[Proposal] Rename `l1_coefficient` to `sparsity_coefficient`
#360
chanind
opened
2 weeks ago
1
feat: adding a CLI training runner
#359
chanind
closed
2 weeks ago
2
Update training_a_sparse_autoencoder.ipynb
#358
ggggggrv15
closed
2 weeks ago
0
chore: more flexible training losses
#357
chanind
closed
2 weeks ago
2
feat: Add linear probe trainer
#356
tom-pollak
opened
3 weeks ago
3
fix: normalize decoder bias in fold_norm_scaling_factor
#355
chanind
closed
3 weeks ago
5
[Bug Report] fold_activation_norm_scaling_factor missing division of decoder bias
#354
tuomaso
closed
3 weeks ago
0
Training time
#353
merlinarer
opened
3 weeks ago
1
Add logic to train JumpReLU SAEs
#352
anthonyduong9
closed
2 weeks ago
5
[Bug Report] get_sae_config doesn't return the same config as SAE.from_pretrained due to defaulting
#351
hijohnnylin
closed
2 weeks ago
2
[Proposal] Allow excluding special tokens when applying SAEs in HookedSAETransformer
#350
chanind
opened
4 weeks ago
1
fix: use the correct layer for new gemma scope SAE sparsities
#349
hijohnnylin
closed
1 month ago
1
fix: use the correct layer for new gemma scope SAE sparsities
#348
hijohnnylin
closed
1 month ago
1
Supply `device` to `SAEConfigLoadOptions`
#347
callummcdougall
closed
1 month ago
2
Add tests for evals
#346
anthonyduong9
closed
1 month ago
0
fix: last NP id fix, hopefully
#345
hijohnnylin
closed
1 month ago
1
fix: np ids should contain model id
#344
hijohnnylin
closed
1 month ago
1
fix: fix duplicate np ids
#343
hijohnnylin
closed
1 month ago
1
fix: yaml was missing some gemmascope np ids, update np id formats
#342
hijohnnylin
closed
1 month ago
0
fix: previous saebench yaml fixes were incomplete for pythia-70m-deduped
#341
hijohnnylin
closed
1 month ago
1
changes dtype default value in read_sae_from_disk()
#340
anthonyduong9
closed
1 month ago
0
updated SAE Bench pythia model names (and loader device cfg)
#339
curt-tigges
closed
1 month ago
1
extracts repeated code to load_pretrained_saes_yaml()
#338
anthonyduong9
opened
1 month ago
0
[Bug Report]
#337
Yoon-Jeong-ho
opened
1 month ago
6
chore: reduce test space usage in CI
#336
chanind
closed
1 month ago
2
[Proposal] Move files in `tests/unit` to `tests`
#335
anthonyduong9
opened
1 month ago
1
Encode with slice
#334
callummcdougall
opened
1 month ago
6
[Proposal] BOS should be ignored by default in the activation store for SAEs not trained on the bos (and future SAEs trained in SAE Lens)
#333
jbloomAus
opened
1 month ago
2
Next