guac commited on
Commit
04653c1
1 Parent(s): a4ee477

[update] Release model weights from deduped Pile training

Browse files
Files changed (3) hide show
  1. README.md +46 -14
  2. config.json +1 -1
  3. pytorch_model.bin +1 -1
README.md CHANGED
@@ -6,15 +6,18 @@ datasets:
6
  - pile
7
  metrics:
8
  - nDCG@10
 
9
  ---
10
 
11
  # Carptriever-1
12
 
13
- # Model description
 
14
 
15
  Carptriever-1 is a `bert-large-uncased` retrieval model trained with contrastive learning via a momentum contrastive (MoCo) mechanism following the work of G. Izacard et al. in ["Contriever: Unsupervised Dense Information Retrieval with Contrastive Learning"](https://arxiv.org/abs/2112.09118).
16
 
17
- # How to use
 
18
 
19
  ```python
20
  from transformers import AutoTokenizer, AutoModel
@@ -51,11 +54,13 @@ for sentence, score in sentence_score_pairs:
51
  print(f"\nSentence: {sentence}\nScore: {score:.4f}")
52
  ```
53
 
 
54
  # Training data
55
 
56
- Carptriever-1 is pre-trained on [The Pile](https://pile.eleuther.ai/), a large and diverse dataset created by EleutherAI for language model training.
57
 
58
- # Training procedure
 
59
 
60
  The model was trained on 32 40GB A100 for approximately 100 hours with the following configurations:
61
 
@@ -73,24 +78,51 @@ The model was trained on 32 40GB A100 for approximately 100 hours with the follo
73
  - `momentum = 0.999`
74
  - `temperature = 0.05`
75
 
76
- # Evaluation results
77
 
78
- We provide evaluation results on the [BEIR: Benchmarking IR](https://github.com/beir-cellar/beir) suite.
 
 
79
 
80
- | nDCG@10 | Avg | MSMARCO | TREC-Covid | NFCorpus | NaturalQuestions | HotpotQA | FiQA | ArguAna | Tóuche-2020 | Quora | CQAdupstack | DBPedia | Scidocs | Fever | Climate-fever | Scifact |
81
- |------------------------------------------|-------------|---------|------------|----------|------------------|----------|------|---------|-------------|-------|-------------|---------|---------|-------|---------------|---------|
82
- | Contriever* | 35.97 | 20.6 | 27.4 | 31.7 | 25.4 | 48.1 | 24.5 | 37.9 | 19.3 | 83.5 | 28.4 | 29.2 | 14.9 | 68.2 | 15.5 | 64.9 |
83
- | Carptriever-1 | 34.29 | 18.81 | **46.5** | 28.9 | 21.1 | 39.01 | 20.2 | 33.4 | 17.3 | 80.6 | 25.4 | 23.6 | 14.9 | 59.6 | **18.7** | **66.4** |
84
 
85
- \* Results are taken from the Contriever [repository](https://github.com/facebookresearch/contriever).
 
 
 
 
 
86
 
87
  Note that degradation in performance, relative to the Contriever model, was expected given the much broader diversity of our training dataset. We plan on addressing this in future updates with architectural improvements and view Carptriever-1 as our first iteration in the exploratory phase towards better language-embedding models.
88
 
89
- # Appreciation
90
 
91
- All compute was graciously provided by [Stability.ai](https://stability.ai/).
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
92
 
93
- # Citations
94
 
95
  ```bibtex
96
  @misc{izacard2021contriever,
 
6
  - pile
7
  metrics:
8
  - nDCG@10
9
+ - MRR
10
  ---
11
 
12
  # Carptriever-1
13
 
14
+
15
+ ## Model description
16
 
17
  Carptriever-1 is a `bert-large-uncased` retrieval model trained with contrastive learning via a momentum contrastive (MoCo) mechanism following the work of G. Izacard et al. in ["Contriever: Unsupervised Dense Information Retrieval with Contrastive Learning"](https://arxiv.org/abs/2112.09118).
18
 
19
+
20
+ ## How to use
21
 
22
  ```python
23
  from transformers import AutoTokenizer, AutoModel
 
54
  print(f"\nSentence: {sentence}\nScore: {score:.4f}")
55
  ```
56
 
57
+
58
  # Training data
59
 
60
+ Carptriever-1 is pre-trained on a de-duplicated subset of [The Pile](https://pile.eleuther.ai/), a large and diverse dataset created by EleutherAI for language model training. This subset was created through a [Minhash LSH](http://ekzhu.com/datasketch/lsh.html) process using a threshold of `0.87`.
61
 
62
+
63
+ ## Training procedure
64
 
65
  The model was trained on 32 40GB A100 for approximately 100 hours with the following configurations:
66
 
 
78
  - `momentum = 0.999`
79
  - `temperature = 0.05`
80
 
 
81
 
82
+ ## Evaluation results
83
+
84
+ #### [BEIR: Benchmarking IR](https://github.com/beir-cellar/beir)
85
 
86
+ We report the following BEIR scores as measured in normalized discounted cumulative gain (nDCG@10):
 
 
 
87
 
88
+ | Model | Avg | MSMARCO | TREC-Covid | NFCorpus | NaturalQuestions | HotpotQA | FiQA | ArguAna | Tóuche-2020 | Quora | CQAdupstack | DBPedia | Scidocs | Fever | Climate-fever | Scifact |
89
+ |---------------|-------|---------|------------|----------|------------------|----------|------|---------|-------------|-------|-------------|---------|---------|-------|---------------|----------|
90
+ | Contriever* | 35.97 | 20.6 | 27.4 | 31.7 | 25.4 | 48.1 | 24.5 | 37.9 | 19.3 | 83.5 | 28.40 | 29.2 | 14.9 | 68.20 | 15.5 | 64.9 |
91
+ | Carptriever-1 | 34.54 | 18.83 | **52.2** | 28.5 | 21.1 | 39.4 | 23.2 | 31.7 | 15.2 | 81.3 | 26.88 | 25.4 | 14.2 | 57.36 | **17.9** | 64.9 |
92
+
93
+ \* Results are taken from the Contriever [GitHub repository](https://github.com/facebookresearch/contriever).
94
 
95
  Note that degradation in performance, relative to the Contriever model, was expected given the much broader diversity of our training dataset. We plan on addressing this in future updates with architectural improvements and view Carptriever-1 as our first iteration in the exploratory phase towards better language-embedding models.
96
 
 
97
 
98
+ #### [CodeSearchNet Challenge Evaluating the State of Semantic Code Search](https://arxiv.org/pdf/1909.09436.pdf)
99
+
100
+ We provide results on the CodeSearchNet benchmark, measured in Mean Reciprocal Rank (MRR), following the code search procedure outlined in Section 3.3 of Neelakantan et al.'s ["Text and Code Embeddings by Contrastive Pre-Training"](https://arxiv.org/pdf/2201.10005.pdf).
101
+
102
+ `Candidate Size = 1,000`
103
+
104
+ | Model | Avg | Python | Go | Ruby | PHP | Java | JS |
105
+ |-----------------|-------|--------|-------|-------|-------|-------|-------|
106
+ | Carptriever-1 | 60.24 | 65.85 | 63.29 | 62.1 | 59.1 | 55.52 | 55.55 |
107
+ | Contriever | 49.39 | 54.81 | 58.9 | 55.19 | 38.46 | 44.89 | 44.09 |
108
+
109
+
110
+ `Candidate Size = 10,000`
111
+
112
+ | Model. | Avg | Python | Go | Ruby | PHP | Java | JS |
113
+ |-----------------|-------|--------|-------|-------|-------|-------|-------|
114
+ | Carptriever-1 | 48.59 | 55.98 | 43.18 | 56.06 | 45.62 | 46.04 | 44.66 |
115
+ | Contriever | 37 | 45.43 | 36.08 | 48.07 | 25.59 | 32.89 | 31.44 |
116
+
117
+
118
+ ## Acknowledgements
119
+
120
+ This work would not have been possible without the compute support of [Stability AI](https://stability.ai/).
121
+
122
+ Thank you to Louis Castricato for research guidance and Reshinth Adithyan for creating the CodeSearchNet evaluation script.
123
+
124
 
125
+ ## Citations
126
 
127
  ```bibtex
128
  @misc{izacard2021contriever,
config.json CHANGED
@@ -20,7 +20,7 @@
20
  "pooling": "average",
21
  "position_embedding_type": "absolute",
22
  "torch_dtype": "float32",
23
- "transformers_version": "4.22.0.dev0",
24
  "type_vocab_size": 2,
25
  "use_cache": true,
26
  "vocab_size": 30522
 
20
  "pooling": "average",
21
  "position_embedding_type": "absolute",
22
  "torch_dtype": "float32",
23
+ "transformers_version": "4.21.3",
24
  "type_vocab_size": 2,
25
  "use_cache": true,
26
  "vocab_size": 30522
pytorch_model.bin CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:dd49ac4881768f3554dd1df5f359421ad807d7643858de0bc06495b541ce8a71
3
  size 1336496437
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:63f84de125685fd11067227632f0be12e124cfde031d4b47460e2220b8ee5f84
3
  size 1336496437