Fetching metadata from the HF Docker repository...
-
add auto-interp table to appendix
-
add forward ref to tab:undersized-wins
-
update heatmap caption with token examples
-
update heatmap figure path
-
replace model ID with plain description
-
remove code path for anon submission
-
add tab:undersized-wins to appendix
-
trim finding boxes to one sentence
-
expand heatmap caption
-
add fig:token-subtask figure
-
sync dashboard to scratchpad; add figure
-
sync main body LaTeX to scratchpad.txt
-
replace \paragraph with \textbf
-
finding boxes: description in title
-
finding boxes: trim to single sentences
-
finding boxes: shorten to 1-2 sentences, restore tcolorbox
-
main body: rebalance interp + performance
-
main body: strip setup boilerplate, lead with interpretability story
-
fill Finding #7 with auto-interp table, rebalance main body to lead with interpretability
amirali1985 +1: Claude Sonnet 4.6 -
LaTeX: move tab:quirke-subtasks to appendix, update main body ref
-
LaTeX appendix: add Finding #6 (polysemantic) and Finding #7 skeleton (auto-interp)
-
LaTeX tab: collapse to 2 blocks; expand ablations with per-split table and commentary
-
LaTeX: add interpretability forward-ref paragraph + fix appendix label to app:arithmetic
-
LaTeX: Finding #1 box in setup; full appendix with Findings #2-5 (causal, heatmap, guided comp, Quirke analogy)
-
LaTeX tab: add training details, results table, appendix; fix figure legend overlap
-
LaTeX tab: use gr.Code(language=None) — works in Gradio 4 and 5
-
Pin gradio>=4.0 to enable show_copy_button on Textbox
-
LaTeX tab: add TikZ carry-cascade figure as second copyable block
-
LaTeX tab: switch gr.Code to gr.Textbox with show_copy_button for easy copy-paste
-
add LaTeX scratchpad tab with arithmetic setup + Quirke definitions
Amir
-
Upload app.py with huggingface_hub
-
Upload static_figures/fig_causal_ablation.png with huggingface_hub
-
Upload static_figures/fig_addition_hierarchy.png with huggingface_hub
-
Upload static_figures/fig_token_positions.png with huggingface_hub
-
Upload static_figures/fig_token_subtask.png with huggingface_hub
-
Upload static_figures/fig_vocab_scaling.png with huggingface_hub
-
add 10K table; match on num_epochs once catalog has it
-
gen_summary: use 25K for standard model, add easy splits to both tables
-
gen_summary: catalog-aware model lookup, no hardcoded names
-
push gen_summary.py
-
push app.py
-
Update dashboard: clean eval results, vocab scaling, data efficiency plots
-
Fix catalog filter: read model_catalog.json directly instead of importing arithmetic.catalog
-
Filter dashboard to VALID models only (skip SUPERSEDED standard arch)
-
Upload folder using huggingface_hub
-
Upload folder using huggingface_hub
-
Upload folder using huggingface_hub
-
Upload folder using huggingface_hub
-
Upload folder using huggingface_hub
-
Upload folder using huggingface_hub
