-
Notifications
You must be signed in to change notification settings - Fork 351
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Difficulty Reproducing BM25 Results #58
Comments
Issue: Discrepancy in Retrieval Results and MetricsDescriptionWe're encountering inconsistencies in the retrieval results that do not match the stated metrics, specifically when evaluating on a dataset with 27,000 contexts at k=50, and across different values of k for a dataset with 13,000 contexts. The discrepancies are evident in the average, all, and any recall metrics, deviating from the expected BM25 recall values. Detailed ObservationsFor 27,000 Contexts at k=50The observed metrics are as follows:
Compared to the expected BM25 recall metrics:
For 13,000 Contexts with Different k ValuesAn increase in metric values is observed with a decrease in k, which is counterintuitive, as the number of files that can be accommodated decreases. At k=10:
At k=50:
At k=3 (Most files are not considered as they have no retrieved files):
Missing Gold Files for Specific CommitsAdditionally, there are instances with missing gold files, indicated by warnings during the retrieval process. Examples include Test Dataset UsedThe test sample for this evaluation is derived from the provided test dataset. RequestWe kindly ask for an investigation into these discrepancies, particularly focusing on:
We believe addressing these points will greatly enhance the accuracy and reliability of the retrieval process, aligning it more closely with the expected outcomes. Thank you for your attention to these matters. |
Tagging @carlosejimenez to address this. |
A relevant question, why recall is a number larger than 1? For example, what does 29.58 mean for 13K, Avg, BM25 Recall. |
@dayuyang1999 Oh I think those are just percentages (29.58%, not an absolute value). We should've put the percentage signs there. |
I've encountered issues while trying to reproduce the BM25 results mentioned in the documentation. I've faced the challenges:
How does the script handle files with more context than the tokenizer can support? Is there a filtering mechanism in place to manage such instances?
Could you provide more details on how the parameter
k
is utilized in the script and its impact on the results?I would appreciate any guidance or suggestions on how to address these issues to achieve the expected BM25 results.
Moreover, the tokenizer is being created for each instance, rather than being kept in memory. This seems to be inefficient and could potentially affect performance.
Also, the tokenization process does not appear to be parallelized. As a result, processing is slow, and when running the test dataset overnight, the scores achieved are lower than expected.
The text was updated successfully, but these errors were encountered: