Document embedding generation details

README.md

@@ -19,7 +19,7 @@ tags:
 
 # Multimodal Embedding 100M
 
-This dataset contains a 100M-row multimodal embedding corpus derived from LAION-400M-style embedding data. It is intended for large-scale vector database ingestion, ANN index construction, nearest-neighbor search, and retrieval benchmark experiments.
+This dataset contains a 100M-row multimodal embedding corpus generated from LAION-400M-style image-text WebDataset shards. It is intended for large-scale vector database ingestion, ANN index construction, nearest-neighbor search, and retrieval benchmark experiments.
 
 The dataset is stored as Parquet files and organized to keep each directory below Hugging Face's per-directory file-count limit.
 
@@ -33,6 +33,65 @@ The dataset is stored as Parquet files and organized to keep each directory belo
 
 Total: 100,020,000 rows across 16,827 Parquet files.
 
+## Embedding Generation
+
+The embeddings were generated from image-caption pairs in the `laion-400m-256` WebDataset shard format. Each input sample was decoded as a PIL image plus caption text, then encoded into one vector with `Qwen/Qwen3-VL-Embedding-8B`.
+
+Generation configuration:
+
+| Property | Value |
+| --- | --- |
+| Embedding model | `Qwen/Qwen3-VL-Embedding-8B` |
+| Inference engine | vLLM, `runner="pooling"` |
+| Inference dtype | `bfloat16` |
+| Batch size | 16 samples per embedding batch |
+| Output dimension | 4096 |
+| Stored dtype | `float32` in Parquet |
+| Normalization | L2 normalized after model inference |
+| Search metric | Inner Product; because vectors are L2-normalized, IP ranking is equivalent to cosine-similarity ranking |
+
+Prompt template:
+
+```python
+[
+    {
+        "role": "system",
+        "content": [{"type": "text", "text": "Represent the user's input."}],
+    },
+    {
+        "role": "user",
+        "content": [
+            {"type": "image", "image": image},
+            {"type": "text", "text": caption},
+        ],
+    },
+]
+```
+
+The prompt was rendered with the model tokenizer's chat template and passed to vLLM together with `multi_modal_data={"image": image}` through the embedding API. Each full image-caption pair produces exactly one 4096-dimensional vector; there is no text chunking or multi-vector output per sample.
+
+Source-field handling:
+
+- Caption text is read from the WebDataset `txt` field when present, otherwise from `json.caption`.
+- Sample key is read from `json.key` when present, otherwise from WebDataset `__key__`.
+- If image preparation fails, the sample falls back to text-only embedding using the same system instruction.
+- If a batch embedding call fails, the batch is retried one sample at a time.
+- If a single-sample embedding call still fails, a zero vector is emitted for that sample before normalization-safe storage.
+
+Split assignment:
+
+- Raw shard outputs are first written as `key:string, emb:fixed_size_list<float32>[4096]`.
+- Final `train` and `test` files add continuous zero-based `id:int64` values.
+- `train` contains the first 100,000,000 rows in deterministic shard order.
+- `test` contains the following 10,000 rows and is disjoint from `train`.
+- The Hugging Face `train/shard-*` directories are only repository-layout shards; they do not change vector ids or row contents.
+
+Ground-truth generation:
+
+- The `neighbors` split contains brute-force top-100 nearest neighbors for all 10,000 test vectors against the 100,000,000 train vectors.
+- Scores are inner products over the L2-normalized vectors.
+- `neighbors[i]` and `scores[i]` are aligned and sorted by descending score.
+
 ## Directory Layout
 
 ```text