Create Adaptive-RETRO-GPT-1B retrieval datastore

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+chunk_metadata.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,22 @@
+---
+tags:
+- retrieval
+- datastore
+- language-modeling
+---
+
+# Adaptive-RETRO-GPT-1B Retrieval Datastore
+
+External chunk datastore used by Adaptive-RETRO-GPT-1B during retrieval-pretraining.
+
+- Source: `wikimedia/wikipedia` / `20231101.en`
+- Chunks: 120000
+- Chunk token window: 512
+- Retriever: signed token-hash inner product top-k
+
+Files:
+
+- `chunks.npy`: tokenized retrieval chunks
+- `embeddings.npy`: normalized chunk embeddings
+- `chunk_metadata.jsonl`: decoded chunk text and metadata
+- `datastore_meta.json`: preprocessing and retrieval config

chunk_metadata.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d5fdfd6bca020acd9f044da752a83995287efe44503d13d5782ce8a8f2ad046
+size 288730554

chunks.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bb84e578aac11b168bbccfa3bba994fd823237b155309312ae5298a3aeac4f84
+size 245760128

datastore_meta.json

@@ -0,0 +1,11 @@
+{
+  "source_dataset": "wikimedia/wikipedia",
+  "source_config": "20231101.en",
+  "docs_seen": 2441,
+  "num_chunks": 120000,
+  "chunk_tokens": 96,
+  "retrieval_seq_len": 512,
+  "hash_dim": 1024,
+  "embedder": "signed_token_hash",
+  "retriever": "exact_inner_product_over_hash_embeddings"
+}

embeddings.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a0ef3df9566e399dfbe1770e98571dc3f31abfb9bb39d1582bbc4437250207ef
+size 245760128