issues
search
understanding-search
/
maze-transformer
This repo is built to facilitate the training and analysis of autoregressive transformers on maze-solving tasks.
24
stars
6
forks
source link
issues
Newest
Newest
Most commented
Recently updated
Oldest
Least commented
Least recently updated
better tests for weight loading with refactoring
#219
mivanit
closed
1 month ago
0
general dependency housekeeping
#218
mivanit
closed
2 months ago
1
Add position embedding type to BaseGPTConfig
#217
afspies
closed
2 months ago
1
Make a way to run tests with a user-specified branch of maze-dataset
#216
aaron-sandoval
closed
2 months ago
1
allow HookedTransformer without ZanjHookedTransformer
#215
junyuanjyjy
closed
3 months ago
0
Support maze dataset tokenizers update
#214
aaron-sandoval
opened
4 months ago
3
update maze-dataset dep and poetry lockfile
#213
mivanit
closed
4 months ago
1
Remove unk token
#212
aaron-sandoval
closed
1 month ago
2
Add check for loading old models
#211
mivanit
opened
6 months ago
0
clean up notebooks
#210
mivanit
opened
7 months ago
0
Fix tokenizer to allow update to `transformers >=4.34.0`
#209
aaron-sandoval
closed
7 months ago
2
Try and fix tokenizer to work with new version of transformers library
#208
afspies
closed
7 months ago
7
Integration of fixes to linting, wandb api usage, readme
#207
mivanit
closed
2 months ago
2
Ran black v24.1.0 on maze_transformers/ and tests/
#206
naveenarun
closed
8 months ago
0
Linter version (Black)
#205
naveenarun
closed
7 months ago
0
Allow training notebook to run without logger or wandb config
#204
naveenarun
closed
8 months ago
1
Add conda instructions to README.md
#203
naveenarun
closed
8 months ago
0
stop integration tests from logging to wandb
#202
mivanit
opened
9 months ago
1
Modify configs to allow specification of pos encoding type
#201
afspies
closed
2 months ago
0
streamline evals
#200
mivanit
opened
10 months ago
0
rework tokenizer to be compatible with huggingface transformers `>=4.34.0`
#199
mivanit
closed
7 months ago
1
minor fixes: loading models from wandb, update some deps
#198
mivanit
closed
1 year ago
1
Add experiments, part 2
#197
mivanit
closed
10 months ago
1
investigate padding issues
#196
mivanit
opened
1 year ago
0
fix broken padding setting in HuggingMazeTokenizer
#195
mivanit
closed
1 year ago
1
allow passing dataset to `train_model`
#194
mivanit
opened
1 year ago
0
Add experiments notebooks
#193
mivanit
closed
1 year ago
1
add experiment notebooks
#192
mivanit
closed
1 year ago
0
Refactor tokenization
#191
mivanit
closed
1 year ago
1
Fix dataset split, pt to zanj model conversion code
#190
mivanit
closed
1 year ago
1
export dataset to external library
#189
mivanit
closed
1 year ago
1
Add optional dataset-only dependencies
#188
mivanit
closed
1 year ago
1
Add optional dataset-only dependencies
#187
mivanit
closed
1 year ago
0
bump muutils and zanj
#186
mivanit
closed
1 year ago
1
Collected Datasets, various fixes
#185
mivanit
closed
1 year ago
2
Constrained dfs, dataset modifications
#184
canrager
closed
1 year ago
4
Constrained depth first search
#183
canrager
closed
1 year ago
1
Zanj datasets getitem
#182
valedan
closed
1 year ago
1
training loop evals
#181
valedan
closed
1 year ago
1
Training batches are missing a bunch of tokens
#180
valedan
closed
1 year ago
2
Enable easier model training without using script
#179
valedan
closed
1 year ago
1
Check for updates on import and warn if update is needed
#178
valedan
closed
1 year ago
1
Zanj integration: datasets & training
#177
mivanit
closed
1 year ago
2
Refactor LatticeMaze.find_shortest_path and fix minor bug
#176
rusheb
opened
1 year ago
4
Calculate appropriate max sequence length based on the dataset
#175
valedan
opened
1 year ago
1
Allow dataset generation to take a seed
#174
valedan
opened
1 year ago
3
Facilitate easy division of dataset into training and validation sets
#173
valedan
closed
1 year ago
3
Automate direct logit attribution process
#172
valedan
opened
1 year ago
2
Show baseline performance in W&B
#171
valedan
opened
1 year ago
0
Create combo dataset
#170
valedan
closed
1 year ago
2
Next