Skip to content

Commit

Permalink
Refactor tolerance settings for MS MARCO dense vector regressions (#2541
Browse files Browse the repository at this point in the history 
)

Continuation of #2538 - refactor tolerance values for HNSW indexes, calibrate wrt flat index scores.
  • Loading branch information
@lintool
lintool committed Jul 8, 2024
1 parent 5eb46b9 commit 3885b5c
Show file tree
Hide file tree
Showing 109 changed files with 616 additions and 490 deletions.
13 changes: 7 additions & 6 deletions docs/regressions/regressions-dl19-passage.bge-base-en-v1.5.hnsw-int8.cached.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -101,16 +101,17 @@ With the above commands, you should be able to reproduce the following results:

| **AP@1000** | **BGE-base-en-v1.5**|
|:-------------------------------------------------------------------------------------------------------------|-----------|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.443 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.444 |
| **nDCG@10** | **BGE-base-en-v1.5**|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.708 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.706 |
| **R@100** | **BGE-base-en-v1.5**|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.614 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.617 |
| **R@1000** | **BGE-base-en-v1.5**|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.843 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.847 |

Note that due to the non-deterministic nature of HNSW indexing, results may differ slightly between each experimental run.
Nevertheless, scores are generally within 0.005 of the reference values recorded in [our YAML configuration file](../../src/main/resources/regression/dl19-passage.bge-base-en-v1.5.hnsw-int8.cached.yaml).
The above figures are from running brute-force search with cached queries on non-quantized **flat** indexes.
With cached queries on quantized HNSW indexes, observed results are likely to differ; scores may be lower by up to 0.01, sometimes more.
Note that both HNSW indexing and quantization are non-deterministic (i.e., results may differ slightly between trials).

❗ Retrieval metrics here are computed to depth 1000 hits per query (as opposed to 100 hits per query for document ranking).
For computing nDCG, remember that we keep qrels of _all_ relevance grades, whereas for other metrics (e.g., AP), relevance grade 1 is considered not relevant (i.e., use the `-l 2` option in `trec_eval`).
Expand Down
11 changes: 6 additions & 5 deletions docs/regressions/regressions-dl19-passage.bge-base-en-v1.5.hnsw-int8.onnx.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -103,14 +103,15 @@ With the above commands, you should be able to reproduce the following results:
|:-------------------------------------------------------------------------------------------------------------|-----------|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.444 |
| **nDCG@10** | **BGE-base-en-v1.5**|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.702 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.706 |
| **R@100** | **BGE-base-en-v1.5**|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.609 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.617 |
| **R@1000** | **BGE-base-en-v1.5**|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.836 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.847 |

Note that due to the non-deterministic nature of HNSW indexing, results may differ slightly between each experimental run.
Nevertheless, scores are generally within 0.005 of the reference values recorded in [our YAML configuration file](../../src/main/resources/regression/dl19-passage.bge-base-en-v1.5.hnsw-int8.onnx.yaml).
The above figures are from running brute-force search with cached queries on non-quantized **flat** indexes.
With ONNX query encoding on quantized HNSW indexes, observed results are likely to differ; scores may be lower by up to 0.01, sometimes more.
Note that both HNSW indexing and quantization are non-deterministic (i.e., results may differ slightly between trials).

❗ Retrieval metrics here are computed to depth 1000 hits per query (as opposed to 100 hits per query for document ranking).
For computing nDCG, remember that we keep qrels of _all_ relevance grades, whereas for other metrics (e.g., AP), relevance grade 1 is considered not relevant (i.e., use the `-l 2` option in `trec_eval`).
Expand Down
11 changes: 6 additions & 5 deletions docs/regressions/regressions-dl19-passage.bge-base-en-v1.5.hnsw.cached.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -99,16 +99,17 @@ With the above commands, you should be able to reproduce the following results:

| **AP@1000** | **BGE-base-en-v1.5**|
|:-------------------------------------------------------------------------------------------------------------|-----------|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.442 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.444 |
| **nDCG@10** | **BGE-base-en-v1.5**|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.706 |
| **R@100** | **BGE-base-en-v1.5**|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.616 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.617 |
| **R@1000** | **BGE-base-en-v1.5**|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.842 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.847 |

Note that due to the non-deterministic nature of HNSW indexing, results may differ slightly between each experimental run.
Nevertheless, scores are generally within 0.005 of the reference values recorded in [our YAML configuration file](../../src/main/resources/regression/dl19-passage.bge-base-en-v1.5.hnsw.cached.yaml).
The above figures are from running brute-force search with cached queries on non-quantized **flat** indexes.
With cached queries on non-quantized HNSW indexes, observed results are likely to differ; scores may be lower by up to 0.01, sometimes more.
Note that HNSW indexing is non-deterministic (i.e., results may differ slightly between trials).

❗ Retrieval metrics here are computed to depth 1000 hits per query (as opposed to 100 hits per query for document ranking).
For computing nDCG, remember that we keep qrels of _all_ relevance grades, whereas for other metrics (e.g., AP), relevance grade 1 is considered not relevant (i.e., use the `-l 2` option in `trec_eval`).
Expand Down
13 changes: 7 additions & 6 deletions docs/regressions/regressions-dl19-passage.bge-base-en-v1.5.hnsw.onnx.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -99,16 +99,17 @@ With the above commands, you should be able to reproduce the following results:

| **AP@1000** | **BGE-base-en-v1.5**|
|:-------------------------------------------------------------------------------------------------------------|-----------|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.447 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.444 |
| **nDCG@10** | **BGE-base-en-v1.5**|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.701 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.706 |
| **R@100** | **BGE-base-en-v1.5**|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.607 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.617 |
| **R@1000** | **BGE-base-en-v1.5**|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.837 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.847 |

Note that due to the non-deterministic nature of HNSW indexing, results may differ slightly between each experimental run.
Nevertheless, scores are generally within 0.005 of the reference values recorded in [our YAML configuration file](../../src/main/resources/regression/dl19-passage.bge-base-en-v1.5.hnsw.onnx.yaml).
The above figures are from running brute-force search with cached queries on non-quantized **flat** indexes.
With ONNX query encoding on non-quantized HNSW indexes, observed results are likely to differ; scores may be lower by up to 0.01, sometimes more.
Note that HNSW indexing is non-deterministic (i.e., results may differ slightly between trials).

❗ Retrieval metrics here are computed to depth 1000 hits per query (as opposed to 100 hits per query for document ranking).
For computing nDCG, remember that we keep qrels of _all_ relevance grades, whereas for other metrics (e.g., AP), relevance grade 1 is considered not relevant (i.e., use the `-l 2` option in `trec_eval`).
Expand Down
13 changes: 7 additions & 6 deletions ...ressions/regressions-dl19-passage.cohere-embed-english-v3.0.hnsw-int8.cached.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -94,16 +94,17 @@ With the above commands, you should be able to reproduce the following results:

| **AP@1000** | **cohere-embed-english-v3.0**|
|:-------------------------------------------------------------------------------------------------------------|-----------|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.487 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.488 |
| **nDCG@10** | **cohere-embed-english-v3.0**|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.690 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.696 |
| **R@100** | **cohere-embed-english-v3.0**|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.647 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.648 |
| **R@1000** | **cohere-embed-english-v3.0**|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.850 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.863 |

Note that due to the non-deterministic nature of HNSW indexing, results may differ slightly between each experimental run.
Nevertheless, scores are generally within 0.005 of the reference values recorded in [our YAML configuration file](../../src/main/resources/regression/dl19-passage.cohere-embed-english-v3.0.hnsw-int8.cached.yaml).
The above figures are from running brute-force search with cached queries on non-quantized **flat** indexes.
With cached queries on quantized HNSW indexes, observed results are likely to differ; scores may be lower by up to 0.01, sometimes more.
Note that both HNSW indexing and quantization are non-deterministic (i.e., results may differ slightly between trials).

❗ Retrieval metrics here are computed to depth 1000 hits per query (as opposed to 100 hits per query for document ranking).
For computing nDCG, remember that we keep qrels of _all_ relevance grades, whereas for other metrics (e.g., AP), relevance grade 1 is considered not relevant (i.e., use the `-l 2` option in `trec_eval`).
Expand Down
13 changes: 7 additions & 6 deletions docs/regressions/regressions-dl19-passage.cohere-embed-english-v3.0.hnsw.cached.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -94,16 +94,17 @@ With the above commands, you should be able to reproduce the following results:

| **AP@1000** | **cohere-embed-english-v3.0**|
|:-------------------------------------------------------------------------------------------------------------|-----------|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.486 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.488 |
| **nDCG@10** | **cohere-embed-english-v3.0**|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.690 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.696 |
| **R@100** | **cohere-embed-english-v3.0**|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.645 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.648 |
| **R@1000** | **cohere-embed-english-v3.0**|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.851 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.863 |

Note that due to the non-deterministic nature of HNSW indexing, results may differ slightly between each experimental run.
Nevertheless, scores are generally within 0.005 of the reference values recorded in [our YAML configuration file](../../src/main/resources/regression/dl19-passage.cohere-embed-english-v3.0.hnsw.cached.yaml).
The above figures are from running brute-force search with cached queries on non-quantized **flat** indexes.
With cached queries on non-quantized HNSW indexes, observed results are likely to differ; scores may be lower by up to 0.01, sometimes more.
Note that HNSW indexing is non-deterministic (i.e., results may differ slightly between trials).

❗ Retrieval metrics here are computed to depth 1000 hits per query (as opposed to 100 hits per query for document ranking).
For computing nDCG, remember that we keep qrels of _all_ relevance grades, whereas for other metrics (e.g., AP), relevance grade 1 is considered not relevant (i.e., use the `-l 2` option in `trec_eval`).
Expand Down
13 changes: 7 additions & 6 deletions docs/regressions/regressions-dl19-passage.cos-dpr-distil.hnsw-int8.cached.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -101,16 +101,17 @@ With the above commands, you should be able to reproduce the following results:

| **AP@1000** | **cosDPR-distil**|
|:-------------------------------------------------------------------------------------------------------------|-----------|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.458 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.466 |
| **nDCG@10** | **cosDPR-distil**|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.717 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.725 |
| **R@100** | **cosDPR-distil**|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.605 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.617 |
| **R@1000** | **cosDPR-distil**|
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.805 |
| [DL19 (Passage)](https://trec.nist.gov/data/deep2020.html) | 0.820 |

Note that due to the non-deterministic nature of HNSW indexing, results may differ slightly between each experimental run.
Nevertheless, scores are generally within 0.005 of the reference values recorded in [our YAML configuration file](../../src/main/resources/regression/dl19-passage.cos-dpr-distil.hnsw-int8.cached.yaml).
The above figures are from running brute-force search with cached queries on non-quantized **flat** indexes.
With cached queries on quantized HNSW indexes, observed results are likely to differ; scores may be lower by up to 0.01, sometimes more.
Note that both HNSW indexing and quantization are non-deterministic (i.e., results may differ slightly between trials).

❗ Retrieval metrics here are computed to depth 1000 hits per query (as opposed to 100 hits per query for document ranking).
For computing nDCG, remember that we keep qrels of _all_ relevance grades, whereas for other metrics (e.g., AP), relevance grade 1 is considered not relevant (i.e., use the `-l 2` option in `trec_eval`).
Expand Down
Loading

0 comments on commit 3885b5c

Please sign in to comment.