jbloomAus DecisionTransformerInterpretability issues

jbloomAus / DecisionTransformerInterpretability

Interpreting how transformers simulate agents performing RL tasks

https://jbloomaus-decisiontransformerinterpretability-app-4edcnc.streamlit.app/

MIT License

58 stars 15 forks source link

issues

Newest

Newest Most commented Recently updated Oldest Least commented Least recently updated

Over resource limits on Streamlit Cloud

#109 eggsyntax closed 3 weeks ago
1
Over resource limits on Streamlit Cloud

#108 mycpuorg opened 3 months ago
0
Over resource limits on Streamlit Cloud

#107 hamzaali98 opened 4 months ago
0
Cuda cannot be disabled

#106 jackmiller2003 opened 8 months ago
3
Major progress on Distill article

#105 JayBaileyCS closed 9 months ago
0
Added Distill scaffolding up to end of introductory section, plus an example doc.

#104 JayBaileyCS closed 9 months ago
0
Created new PR for pull 54. Added extra test as requested.

#103 JayBaileyCS opened 9 months ago
2
Swapped device to str instead of torch.device to fix PyArrow crash problem

#102 JayBaileyCS closed 9 months ago
0
Added save function to save embedding directions.

#101 JayBaileyCS closed 9 months ago
0
Added text to cosine embedding heatmap of size 10 or less. Increased font size.

#100 JayBaileyCS closed 9 months ago
0
Added visualisation improvements for post

#99 JayBaileyCS closed 9 months ago
1
Added attention pattern and logit lens analyses for patched activations

#98 JayBaileyCS closed 9 months ago
0
Added light mode friendly checkbox to Embeddings for PCA graphs

#97 JayBaileyCS closed 9 months ago
0
Initial RTG slider change and minor others

#96 Mjahaha closed 9 months ago
0
Added text annotations to congruence gridmaps

#95 JayBaileyCS closed 9 months ago
0
Replaced patching coordinates with single searchbox with absolute coords

#94 JayBaileyCS closed 9 months ago
0
Changing the default of Positive Action Direction from Forward to Left

#93 Mjahaha closed 9 months ago
0
Fixed error with recent gridmaps

#92 JayBaileyCS closed 9 months ago
1
changed error in path patching for same tokens to warning

#91 Mjahaha closed 9 months ago
0
Fixed tuple error in path patching

#90 JayBaileyCS closed 10 months ago
0
Added ability to select by one or more states to Embeddings tab.

#89 JayBaileyCS closed 10 months ago
1
Added cache to most static analyses, refactored app.py.

#88 JayBaileyCS closed 10 months ago
0
Added Attention Patterns by RTG tab.

#87 JayBaileyCS closed 10 months ago
0
Added Docker and profiling instructions

#86 JayBaileyCS closed 10 months ago
1
Improved dockerfile and added trajectory collection code to evaluate_dt_agent

#85 JayBaileyCS closed 11 months ago
1
Improved dockerfile and added dockerignore to speed up Docker build

#84 JayBaileyCS closed 11 months ago
0
Fixed ablation tests

#83 JayBaileyCS closed 1 year ago
0
"Algebraic value editing" raises exception

#82 alexander-turner opened 1 year ago
1
Complete QK/OV Circuit visualizations

#81 jbloomAus opened 1 year ago
0
Fix Ablation Tool

#80 jbloomAus opened 1 year ago
0
Shapley Values on Attention Heads or Causal Edges Via Ablation

#79 jbloomAus opened 1 year ago
0
Complete Embedding visualizations

#78 jbloomAus opened 1 year ago
0
Reverse Logit Lense

#77 jbloomAus opened 1 year ago
0
Look into why MemoryDT appears to have no bias on the value terms.

#76 jbloomAus closed 1 year ago
1
Expand analytical AVEC

#75 jbloomAus opened 1 year ago
0
Write a post before EAG London

#74 jbloomAus closed 1 year ago
2
Streamlit app requires mujoco installation

#73 DalasNoin opened 1 year ago
0
Implement AVEC in the interpretability app

#72 jbloomAus closed 1 year ago
2
Folding Layer Norm in Model Loading

#71 jbloomAus closed 1 year ago
9
Make it possible to track the preferences of the PPO in the app.

#70 jbloomAus opened 1 year ago
1
SVD Decomp / Explore ways to use dimensionality reduction to quickly understand what heads are doing.

#69 jbloomAus opened 1 year ago
1
Improve history panel in streamlit app

#68 jbloomAus closed 1 year ago
1
Facelift of the RTG Scan in the streamlit app

#67 jbloomAus closed 1 year ago
1
Train a BC on PCT traj = 1 with two different agents mixed in and see if we can tell which one it thinks it is.

#66 jbloomAus opened 1 year ago
0
Verify Initialization of Transformer Model Components is good/appropriate.

#65 jbloomAus closed 1 year ago
6
Add model export option using ONNX to facilitate better Netron visualization

#64 jbloomAus opened 1 year ago
0
Write a check to look at layer weight norms at initialization on the architecture, maybe visualize in a bar chart.

#63 jbloomAus closed 1 year ago
0
Check how LSTM model BOW init is being done and whether it needs a fix

#62 jbloomAus opened 1 year ago
0
Better encode/embed MiniGrid State to speed up training in DT's.

#61 jbloomAus closed 1 year ago
7
Add static interpretability visualizations to wandb dashboard.

#60 jbloomAus opened 1 year ago
0