Skip to content

pgiouroukis/anyedit

Repository files navigation

Investigating Long-Form Knowledge Editing Across Small Language Models

Final Project implementation for Columbia COMS4705

The following commands were used to run the experiments for the UnKEBench dataset. In each case, the first command runs the AnyEdit algorithm with the MEMIT backbone (as they do in the AnyEdit paper) on the specified model and hyperparameters file. We set the dataset size limit to 10 for quick evaluation. We only perform 1 edit per run, following the AnyEdit paper. The second command computes and prints the metrics for the quality of the generation. The following metrics are computed: Bleu, Rouge-1, Rouge-2, Rouge-L, and BERTScore.

We used an A100 GPU with 80GB for running these experiments, since the MEMIT backbone requires significant GPU memory.

Gemma3-1B-it

python3 -m experiments.evaluate_uns --alg_name=MEMIT_ARE --model_name=google/gemma-3-1b-it --hparams_fname=Gemma3-1B-it.json --ds_name=unke --dataset_size_limit=10 --num_edits=1

python -m experiments.summarize_uns --file_path=/home/petros/msc/nlp/project/AnyEdit/output/MEMIT_ARE_google_gemma-3-1b-it_unke_result.json

Gemma3-4B-it

python3 -m experiments.evaluate_uns --alg_name=MEMIT_ARE --model_name=google/gemma-3-4b-it --hparams_fname=Gemma3-4B-it.json --ds_name=unke --dataset_size_limit=10 --num_edits=1

python -m experiments.summarize_uns --file_path=/home/petros/msc/nlp/project/AnyEdit/output/MEMIT_ARE_google_gemma-3-4b-it_unke_result.json

Qwen2.5-3B-Instruct

python3 -m experiments.evaluate_uns --alg_name=MEMIT_ARE --model_name=Qwen/Qwen2.5-3B-Instruct --hparams_fname=Qwen2.5-3B-Instruct.json --ds_name=unke --dataset_size_limit=10 --num_edits=1

python -m experiments.summarize_uns --file_path=/home/petros/msc/nlp/project/AnyEdit/output/MEMIT_ARE_Qwen2.5-3B-Instruct_unke_result.json

Lamma 3.2-3B

python3 -m experiments.evaluate_uns --alg_name=MEMIT_ARE --model_name=meta-llama/Llama-3.2-3B --hparams_fname=Llama3.2-3B.json --ds_name=unke --dataset_size_limit=10 --num_edits=1

python -m experiments.summarize_uns --file_path=/home/petros/msc/nlp/project/AnyEdit/output/MEMIT_ARE_meta-llama_Llama-3.2-3B_unke_result.json

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors