Import model checkpoints from public URLs (CCRI-POPROX#106)

This updates our model checkpoint `.dvc` files to reference public(-ish)
URLs to allow external researchers to more easily obtain the data to use
`poprox-recommender`.

- HuggingFace models are directly imported from HuggingFace.
- Our model checkpoints are now imported from an S3 bucket for public
checkpoints (which is not yet actually publicly-accessible, but we will
fix that).
- MIND data must still be obtained from Microsoft and manually added.

It also removes the unused `word2int.tsv` and `ckpt-30000.pth` files.

Eventually, we can update the model checkpoints to import from the
training DVC repository.

models/distilbert-base-uncased/model.safetensors.dvc

@@ -3,3 +3,10 @@ outs:
   size: 267954768
   hash: md5
   path: model.safetensors
+md5: dd731a6d839cbfdd52ba4e065b61f265
+frozen: true
+deps:
+- path: model.safetensors
+  repo:
+    url: https://huggingface.co/distilbert/distilbert-base-uncased
+    rev_lock: 12040accade4e8a0f71eabdb258fecc2e7e948be

models/distilbert-base-uncased/tokenizer.json.dvc

@@ -3,3 +3,10 @@ outs:
   size: 466062
   hash: md5
   path: tokenizer.json
+md5: d2bcada8b5513c9210ba8928bafd2124
+frozen: true
+deps:
+- path: tokenizer.json
+  repo:
+    url: https://huggingface.co/distilbert/distilbert-base-uncased
+    rev_lock: 12040accade4e8a0f71eabdb258fecc2e7e948be

models/model.safetensors.dvc

@@ -3,3 +3,10 @@ outs:
   size: 142805680
   hash: md5
   path: model.safetensors
+md5: 7e0fb2f50423e59a2a62a4d660297f80
+frozen: true
+deps:
+- etag: 851304fad87aa595b26edf79b1a606e0-18
+  size: 142805680
+  hash: md5
+  path: s3://poprox-public-models/nrms-baseline/model.safetensors

models/news-category-classifier-distilbert/model.safetensors.dvc

@@ -3,3 +3,10 @@ outs:
   size: 438081688
   hash: md5
   path: model.safetensors
+md5: b7c02937f341f2e16506b29425e280be
+frozen: true
+deps:
+- path: model.safetensors
+  repo:
+    url: https://huggingface.co/dima806/news-category-classifier-distilbert
+    rev_lock: e6d8809b863f80b7cd2bbaf241112b0b926e0da8

models/news-category-classifier-distilbert/vocab.txt.dvc

@@ -3,3 +3,10 @@ outs:
   size: 231508
   hash: md5
   path: vocab.txt
+md5: 721f828a6f53b9b76a86fbdcbdea2617
+frozen: true
+deps:
+- path: vocab.txt
+  repo:
+    url: https://huggingface.co/dima806/news-category-classifier-distilbert
+    rev_lock: e6d8809b863f80b7cd2bbaf241112b0b926e0da8

models/news_encoder.safetensors.dvc

@@ -3,3 +3,10 @@ outs:
   size: 275529184
   hash: md5
   path: news_encoder.safetensors
+md5: 0c2738f5293159e8e95013eb32c8024a
+frozen: true
+deps:
+- etag: 3de466cec52fde652d1642f55d3bb2f5-33
+  size: 275529184
+  hash: md5
+  path: s3://poprox-public-models/nrms-baseline/news_encoder.safetensors

models/user_encoder.safetensors.dvc

@@ -3,3 +3,10 @@ outs:
   size: 10066184
   hash: md5
   path: user_encoder.safetensors
+md5: cab8a0941111aa717a2a69011851d401
+frozen: true
+deps:
+- etag: c8dcb260de4589573a8aa693466fff79-2
+  size: 10066184
+  hash: md5
+  path: s3://poprox-public-models/nrms-baseline/user_encoder.safetensors