goal: update plotting functions in e2e_analyze.py to output the format used in the paper
example

• allow reducing verbosity
• add progress bar to long-running loops
• provide start and stop timestamps on most important functions
• potentially: log basic info about the machine we're running on

#47 adds barplots grouped by dataset. These plots are still missing errorbars. Asymmetric errorbars for pandas barplots turn out to be surprisingly tricky to get right; hence, the separate issue

We want to see (tentatively) whether model edits are more, less, or equally sensitive if the model size increases

Goal: confirm that specificity++ works by

• running it on the models for which specificity+ results are already available
• evaluating the results
• comparing to the specificity+ results

Task Description: Ease-of-use enhancement allowing end-to-end experiments with a single call

Background

Our experiments require the outputs (incl. next token log probs) of the unedited model and multiple edit algorithms. Currently, the evaluation is a separate step and requires a separate call for each algorithm, leading to inefficiencies in computation and mental overhead.

Goal

To increase efficiency and ease of handling, we aim to evaluate a given test dataset batch directly for all edit algorithms and the unedited model. This dispenses with the need for transferring log prob files between the compute node and DFS when running on a cluster, and makes it easier to quickly experiment with all edit algorithms on a small slice of the test data.

Tasks

• Chain the runs of multiple edit algorithms and the unedited model on a given test dataset batch
• Perform the evaluation of the test dataset batch
• adapt analysis and plotting script to combine outputs for several test dataset batches

Clean up the README file

Currently, experiments/evaluate.py does not store info about which LLM has been anywhere (see screenshot below for an overview over which information is available, currently)

Goal: Make sure the model which has been used gets logged inside the run_dir. Best will be to log it in the *case_*.json file since then it would be traceable on the individual test case level (in principle, edits for different LLMs could end up in the same run_dir if the --continue_from_run flag is used)

The reporting of experimental results in summarize.py should be improved to facilitate a more in-depth analysis:

• add reporting to also include case ID etc
• make it easier to save evaluation results in reasonable format (JSONL)
• expose all arguments of summarize.main in the CLI
• allow to read reports into pandas DataFrame easily

current state

the logits are slightly different from one run to the next

goal

make it deterministic so it's easier to see when we're introducing actual bugs

Goal: take the output of runs of our e2e.py script and produce statistics and ploys for our paper.

Output of e2e.py: results.csv files in the "combined" subfolder of our results directory.

To do: new script e2e_analyze.py which produces statistics and plots using these four lines from experiments/analyze.py

    dfs = get_statistics(df) 
     print(format_statistics(dfs)) 
     plot_statistics(dfs) 
     export_statistics(dfs)

the code in experiments/evaluate.py behaves differently in the MEMIT repo compared to the ROME repo:

• in the ROME repo, for every run of an evaluation it would run the edited as well as the unedited model on every test case; results for the former would be stored under the "post" key in the case_{case_id}.json file, results for the latter in the "pre" key.
• in the MEMIT repo, it seems to run only the unedited model on every test case. This is more efficient, since we don't need to run the unedited model again for every memory editing algo.

goal: allow to run the unedited model as well using the CLI

• consume results for unedited model from results/IDENTITY/... folder
• use information about the used LLM and editing algo from the results json file if available
• extract results for specificity++
• inspect outliers

• experiments
• cluster_experiments
• layer_stats

Goal

Allow to evaluate specificity on CounterFact with the edit-prompt prepended to the test prompt and using KL divergence between edited and unedited model as the metric.

Details

• build on the code from jas-ho/romepp#1
• make sure we still get all three numbers (specificity, sp+, sp++) at the same time for comparison

Add results to the main text and figure

Esben needs maintainer access to connect the website or we should move it to https://github.com/apartresearch.