constella-zero — FastEmbed-first card, run p35w-2m-s2500
Browse files
README.md
CHANGED
|
@@ -1,273 +1,202 @@
|
|
| 1 |
---
|
| 2 |
license: mit
|
| 3 |
language: en
|
| 4 |
-
library_name:
|
| 5 |
tags:
|
|
|
|
|
|
|
|
|
|
| 6 |
- retrieval
|
| 7 |
-
- sentence-similarity
|
| 8 |
- asymmetric-dual-encoder
|
| 9 |
- edge
|
| 10 |
-
- quantized
|
| 11 |
base_model: NovaSearch/stella_en_400M_v5
|
| 12 |
pipeline_tag: feature-extraction
|
| 13 |
---
|
| 14 |
|
| 15 |
-
# zero
|
| 16 |
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
1024-d). Queries are encoded on the edge by **this table**: 30,522 × 1024 int8 rows and one
|
| 20 |
-
pooling rule. Encoding a query is a gather and a weighted sum — **no transformer, no matmul,
|
| 21 |
-
no GPU**. Sub-millisecond per query on one CPU core; the whole query asset is **31.8 MB**.
|
| 22 |
|
| 23 |
-
|
| 24 |
-
|
|
|
|
| 25 |
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 29 |
|
| 30 |
## Usage
|
| 31 |
|
| 32 |
-
The
|
| 33 |
|
| 34 |
```python
|
| 35 |
-
|
| 36 |
-
from huggingface_hub import snapshot_download
|
| 37 |
-
import sys
|
| 38 |
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
enc = ZeroQueryEncoder(d, variant="int8") # or "fp16"
|
| 44 |
-
q = enc.encode(["how do mrna vaccines work?"]) # (1, 1024), L2-normalized
|
| 45 |
```
|
| 46 |
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
The same query path is also an ONNX graph, so you can serve it without numpy or `tokenizers`
|
| 50 |
-
Python — no transformer, still a gather and a sum. `model.onnx` returns the pooled, normalized
|
| 51 |
-
vector; `model_tokens.onnx` returns one weighted row per token, for pipelines that do their own
|
| 52 |
-
masked-mean pooling (fastembed's, for instance).
|
| 53 |
|
| 54 |
-
|
| 55 |
-
# pip install onnxruntime tokenizers
|
| 56 |
-
import numpy as np, onnxruntime as ort
|
| 57 |
-
from tokenizers import Tokenizer
|
| 58 |
-
|
| 59 |
-
tok = Tokenizer.from_file(f"{d}/tokenizer.json") # padding/truncation already correct
|
| 60 |
-
sess = ort.InferenceSession(f"{d}/model.onnx", providers=["CPUExecutionProvider"])
|
| 61 |
-
|
| 62 |
-
encs = [tok.encode(t) for t in ["how do mrna vaccines work?", "who invented the barometer"]]
|
| 63 |
-
L = max(len(e.ids) for e in encs)
|
| 64 |
-
ids = np.zeros((len(encs), L), np.int64)
|
| 65 |
-
attn = np.zeros((len(encs), L), np.int64)
|
| 66 |
-
for i, e in enumerate(encs):
|
| 67 |
-
ids[i, :len(e.ids)], attn[i, :len(e.ids)] = e.ids, 1
|
| 68 |
-
|
| 69 |
-
Q = sess.run(None, {"input_ids": ids, "attention_mask": attn})[0] # (2, 1024), normalized
|
| 70 |
-
assert Q.shape == (2, 1024)
|
| 71 |
-
```
|
| 72 |
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
rows would cost. **You need exactly one of `model.npz`, `model.onnx` or `model_tokens.onnx`**, not
|
| 76 |
-
all three; the repo ships all of them so you can pick your runtime.
|
| 77 |
|
| 78 |
-
|
| 79 |
-
against this exact document space:
|
| 80 |
|
| 81 |
```python
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
doc_model = SentenceTransformer(
|
| 87 |
-
"NovaSearch/stella_en_400M_v5",
|
| 88 |
-
revision="ffeb2b7ee715c226d4ffe5e4619f7dbb48624c20",
|
| 89 |
-
trust_remote_code=True,
|
| 90 |
-
# required unless xformers is installed; also the pinned setting the table was distilled under
|
| 91 |
-
config_kwargs={"use_memory_efficient_attention": False, "unpad_inputs": False},
|
| 92 |
-
)
|
| 93 |
docs = [
|
| 94 |
"mRNA vaccines deliver a strand of messenger RNA encoding a viral antigen.",
|
| 95 |
"The Treaty of Westphalia ended the Thirty Years' War in 1648.",
|
| 96 |
]
|
| 97 |
-
D = doc_model.
|
| 98 |
-
|
| 99 |
-
scores = q @ D.T
|
| 100 |
-
print(docs[int(np.argmax(scores))])
|
| 101 |
```
|
| 102 |
|
| 103 |
-
That asymmetry is the point: `doc_model`
|
| 104 |
-
every query, on the device, and costs almost nothing.
|
| 105 |
-
|
| 106 |
-
## Using it with Qdrant
|
| 107 |
|
| 108 |
-
|
| 109 |
-
L2-normalized, which means `DOT` ranks identically to `COSINE` and is cheaper.
|
| 110 |
|
| 111 |
```python
|
| 112 |
from qdrant_client import QdrantClient, models
|
| 113 |
|
| 114 |
client = QdrantClient(":memory:") # or your cluster
|
| 115 |
client.create_collection("docs", vectors_config=models.VectorParams(
|
| 116 |
-
size=1024, distance=models.Distance.
|
| 117 |
client.upsert("docs", points=[
|
| 118 |
models.PointStruct(id=i, vector=D[i].tolist(), payload={"text": t})
|
| 119 |
-
for i, t in enumerate(docs)])
|
| 120 |
|
| 121 |
-
hits = client.query_points("docs", query=q
|
| 122 |
-
print(hits[0].payload["text"])
|
| 123 |
```
|
| 124 |
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
indexing it inflated the shard from 466 MB to 1.82 GB for no benefit. The query path becomes
|
| 128 |
-
tokenize → fetch rows by id → pool → search, with no model weights in your process at all.
|
| 129 |
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
not the published one — combine the scores yourself to reproduce it. Dense-only in Qdrant
|
| 134 |
-
reproduces 0.4339 exactly **with exact search** — the HNSW advice above is about cost, and ANN
|
| 135 |
-
recall is a separate confound the published numbers deliberately avoid.
|
| 136 |
|
| 137 |
-
|
| 138 |
-
was never measured for quality here** — treat it as untested. (The int8 quality-free result
|
| 139 |
-
above is about the *query table*, not the document index.)
|
| 140 |
|
| 141 |
-
|
|
|
|
| 142 |
|
| 143 |
-
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
|
| 147 |
-
|
| 148 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 149 |
|
| 150 |
## Files
|
| 151 |
|
| 152 |
-
|
| 153 |
-
|
| 154 |
-
|
|
| 155 |
-
| `model.onnx` | the whole query path as one opset-17 graph → `(b, 1024)` pooled + normalized |
|
| 156 |
-
| `model_tokens.onnx` | the same, emitting `(b, s, 1024)` per-token weighted rows for pipelines that pool themselves |
|
| 157 |
-
| `config.json` | the frozen preprocessing rule, teacher pin, document-encoder spec, shas |
|
| 158 |
-
| `zero_encoder.py` | the whole query path — numpy + tokenizers, no torch |
|
| 159 |
-
| `tokenizer.json`, `vocab.txt`, … | stella's WordPiece tokenizer, copied at the pinned revision — with the two edits below |
|
| 160 |
-
|
| 161 |
-
**Two deliberate edits to the copied tokenizer files.** stella's own files declare
|
| 162 |
-
`model_max_length: 32768`, `max_length: 8000` and fixed-512 padding; the frozen rule here — and
|
| 163 |
-
the document index — are **512 with no padding**. This bundle ships `model_max_length: 512`,
|
| 164 |
-
`max_length: 512` and `padding: null`. `config.json` records the original values under
|
| 165 |
-
`tokenizer_deviation_from_teacher`.
|
| 166 |
-
|
| 167 |
-
Precisely what that changes, by caller:
|
| 168 |
-
|
| 169 |
-
| caller | before | after |
|
| 170 |
|---|---|---|
|
| 171 |
-
| `
|
| 172 |
-
| `
|
| 173 |
-
| `
|
| 174 |
-
| Sentence Transformers | requests truncation itself and may impose its own `max_seq_length` | unchanged in that respect; only the underlying default moves to 512 |
|
| 175 |
-
| fastembed's tokenizer loader | truncation 8000, fixed-512 padding kept | truncation 512, dynamic padding |
|
| 176 |
|
| 177 |
-
|
| 178 |
-
|
| 179 |
-
`fastembed.common.preprocessor_utils.load_tokenizer` only — **this repo ships no ONNX graph**, so
|
| 180 |
-
there is no fastembed inference path to describe yet.
|
| 181 |
|
| 182 |
-
|
| 183 |
-
`
|
| 184 |
-
|
|
|
|
| 185 |
|
| 186 |
## Results
|
| 187 |
|
| 188 |
-
nDCG@10, exact search
|
| 189 |
-
|
| 190 |
-
the system was frozen (sha `a7007b1a…`) before any of them were observed.
|
| 191 |
|
| 192 |
-
| system | arguana | fiqa | nfcorpus | scidocs | scifact | trec-covid | **
|
| 193 |
|---|---|---|---|---|---|---|---|
|
| 194 |
-
| **
|
| 195 |
-
|
|
| 196 |
| BM25 alone | 0.4878 | 0.2532 | 0.3180 | 0.1565 | 0.6791 | 0.6099 | 0.4174 |
|
| 197 |
-
|
|
| 198 |
|
| 199 |
-
|
|
|
|
| 200 |
|
| 201 |
-
|
| 202 |
-
|
| 203 |
-
| `zero` > LightRetriever dense (0.4583) | **−0.0243** | [−0.0405, −0.0086] | **miss, resolved below the bar** |
|
| 204 |
-
| `zero` > BM25 (0.4174) | +0.0165 | [+0.0017, +0.0311] | CI passes, multiplicity does not — unresolved |
|
| 205 |
-
| fused > OpenSearch sparse (0.4868) | +0.0043 | [−0.0063, +0.0151] | **statistical tie** |
|
| 206 |
|
| 207 |
-
|
| 208 |
-
sets at ~1/1000 of the query-side cost, and beats BM25 on average — but it does **not** beat a
|
| 209 |
-
comparable small dense query encoder, and the honest headline is the miss. The one bright spot
|
| 210 |
-
is fusion: `zero` + BM25 is a **statistical tie with OpenSearch's learned sparse retriever**
|
| 211 |
-
while its query side remains a table lookup plus token counts.
|
| 212 |
-
|
| 213 |
-
### Caveats you should read before trusting a number
|
| 214 |
|
| 215 |
- **Teacher contamination.** stella discloses **ArguAna** and **FiQA** in its training data —
|
| 216 |
-
|
| 217 |
-
|
| 218 |
-
|
| 219 |
-
|
| 220 |
-
|
| 221 |
-
|
| 222 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 223 |
|
| 224 |
## Costs
|
| 225 |
|
| 226 |
| | |
|
| 227 |
|---|---|
|
| 228 |
-
| query asset (int8 rows + scales + tokenizer) |
|
| 229 |
-
|
|
| 230 |
-
| `model.onnx`, batch 1, one thread,
|
| 231 |
-
| `
|
| 232 |
-
| hydration (cold load to first query) |
|
| 233 |
-
| document
|
| 234 |
-
| document index, 1024-d int8 | 1.02 GB per 1M documents |
|
| 235 |
-
|
| 236 |
-
The ONNX graph derives token counts from an all-pairs comparison, so its cost grows with the
|
| 237 |
-
SQUARE of the sequence length — 26x from an 8-token query to a 512-token one. Real queries sit at
|
| 238 |
-
the short end (the dev set's median is 13 wordpieces), but a long one is not free.
|
| 239 |
|
| 240 |
-
|
| 241 |
-
|
|
|
|
| 242 |
|
| 243 |
## Training
|
| 244 |
|
| 245 |
-
L2 regression of the table's pooled output onto
|
| 246 |
-
|
| 247 |
-
**
|
| 248 |
|
| 249 |
-
|
| 250 |
-
(
|
| 251 |
-
still misses the bar. The gap is architectural, not licensing.)
|
| 252 |
|
| 253 |
-
##
|
| 254 |
-
|
| 255 |
-
NQ, SQuAD, HotpotQA, FEVER and Mr. TyDi derive from Wikipedia and are **CC BY-SA** (3.0/4.0).
|
| 256 |
-
Amazon ESCI is Apache-2.0. The teacher, `NovaSearch/stella_en_400M_v5`, is MIT.
|
| 257 |
-
|
| 258 |
-
## Revisions
|
| 259 |
-
|
| 260 |
-
| | |
|
| 261 |
-
|---|---|
|
| 262 |
-
| first published | 2026-09-03 — the frozen M7 bundle, with stella's tokenizer files copied verbatim |
|
| 263 |
-
| ONNX added | 2026-09-03, commit `fb8e5c5b` — `model.onnx` and `model_tokens.onnx`; the `.npz` and its numbers unchanged |
|
| 264 |
-
| tokenizer fixed | 2026-09-03, commit `1aa60418` — `tokenizer_config.json` `model_max_length`/`max_length` 32768/8000 → 512, `tokenizer.json` `padding` `Fixed(512)` → `null`, and one broken snippet in this card fixed |
|
| 265 |
-
|
| 266 |
-
`model.npz` is byte-identical across both (sha `a7007b1a…`) and the reference encoder's output is
|
| 267 |
-
unchanged, so **no published number differs between revisions**. Pass `revision=` to
|
| 268 |
-
`snapshot_download` if you need to pin one.
|
| 269 |
-
|
| 270 |
-
## Citation / provenance
|
| 271 |
|
| 272 |
```
|
| 273 |
run_id p35w-2m-s2500
|
|
@@ -276,3 +205,5 @@ teacher NovaSearch/stella_en_400M_v5 @ ffeb2b7ee715c226d4ffe5e4619f7d
|
|
| 276 |
preproc prefix="" · add_special_tokens · max_length=512 · pool_mode=sqrt
|
| 277 |
preproc fingerprint adb24fb2e8cad66f
|
| 278 |
```
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: mit
|
| 3 |
language: en
|
| 4 |
+
library_name: fastembed
|
| 5 |
tags:
|
| 6 |
+
- fastembed
|
| 7 |
+
- qdrant
|
| 8 |
+
- onnx
|
| 9 |
- retrieval
|
|
|
|
| 10 |
- asymmetric-dual-encoder
|
| 11 |
- edge
|
|
|
|
| 12 |
base_model: NovaSearch/stella_en_400M_v5
|
| 13 |
pipeline_tag: feature-extraction
|
| 14 |
---
|
| 15 |
|
| 16 |
+
# constella-zero
|
| 17 |
|
| 18 |
+
The query side of an **asymmetric dual encoder**: documents are indexed once, in the cloud, by a
|
| 19 |
+
large frozen encoder; queries are encoded on the device by **a lookup table**.
|
|
|
|
|
|
|
|
|
|
| 20 |
|
| 21 |
+
There is no transformer here. The model is 30,522 × 1024 int8 rows and one pooling rule —
|
| 22 |
+
encoding a query is a gather and a weighted sum. The query asset is **31.8 MB** and a query costs
|
| 23 |
+
**~0.05 ms** on one CPU core.
|
| 24 |
|
| 25 |
+
It was distilled from [`stella_en_400M_v5`](https://huggingface.co/NovaSearch/stella_en_400M_v5)
|
| 26 |
+
so that its output lands in that model's document space. The matching document encoder is
|
| 27 |
+
published as [`stella-en-400M-v5-doc-onnx`](https://huggingface.co/DylanCouzon/stella-en-400M-v5-doc-onnx);
|
| 28 |
+
the two are only meaningful together.
|
| 29 |
+
|
| 30 |
+
*constella = constellation + stella: navigate by fixed stars, no engine.*
|
| 31 |
+
|
| 32 |
+
> **Research preview.** It is a bag of tokens and behaves like one. Read
|
| 33 |
+
> [Results](#results) and [Limits](#limits) first.
|
| 34 |
|
| 35 |
## Usage
|
| 36 |
|
| 37 |
+
The snippets in this section run in order, sharing state.
|
| 38 |
|
| 39 |
```python
|
| 40 |
+
from fastembed import TextEmbedding
|
|
|
|
|
|
|
| 41 |
|
| 42 |
+
NAME = "DylanCouzon/constella-zero"
|
| 43 |
+
query_model = TextEmbedding(NAME)
|
| 44 |
+
q = next(iter(query_model.embed(["how do mrna vaccines work?"]))) # (1024,), L2-normalized
|
|
|
|
|
|
|
|
|
|
| 45 |
```
|
| 46 |
|
| 47 |
+
Not in a FastEmbed release yet. Until it is:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 48 |
|
| 49 |
+
pip install "fastembed @ git+https://github.com/Dylancouzon/fastembed@add-constella-models"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 50 |
|
| 51 |
+
FastEmbed fetches only `model.onnx` and the tokenizer — about 31 MB, not the whole repo. Pooling
|
| 52 |
+
and L2 normalization happen inside the graph.
|
|
|
|
|
|
|
| 53 |
|
| 54 |
+
### The document side
|
|
|
|
| 55 |
|
| 56 |
```python
|
| 57 |
+
DOC_NAME = "DylanCouzon/stella-en-400M-v5-doc-onnx"
|
| 58 |
+
|
| 59 |
+
doc_model = TextEmbedding(DOC_NAME) # 1.75 GB, runs in the cloud, once per document
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 60 |
docs = [
|
| 61 |
"mRNA vaccines deliver a strand of messenger RNA encoding a viral antigen.",
|
| 62 |
"The Treaty of Westphalia ended the Thirty Years' War in 1648.",
|
| 63 |
]
|
| 64 |
+
D = list(doc_model.embed(docs))
|
|
|
|
|
|
|
|
|
|
| 65 |
```
|
| 66 |
|
| 67 |
+
That asymmetry is the point: `doc_model` is a 400M-parameter transformer that runs once per
|
| 68 |
+
document. `query_model` runs on every query, on the device, and costs almost nothing.
|
|
|
|
|
|
|
| 69 |
|
| 70 |
+
### With Qdrant
|
|
|
|
| 71 |
|
| 72 |
```python
|
| 73 |
from qdrant_client import QdrantClient, models
|
| 74 |
|
| 75 |
client = QdrantClient(":memory:") # or your cluster
|
| 76 |
client.create_collection("docs", vectors_config=models.VectorParams(
|
| 77 |
+
size=1024, distance=models.Distance.COSINE))
|
| 78 |
client.upsert("docs", points=[
|
| 79 |
models.PointStruct(id=i, vector=D[i].tolist(), payload={"text": t})
|
| 80 |
+
for i, t in enumerate(docs)])
|
| 81 |
|
| 82 |
+
hits = client.query_points("docs", query=q.tolist(), limit=5).points
|
| 83 |
+
print(hits[0].payload["text"])
|
| 84 |
```
|
| 85 |
|
| 86 |
+
Qdrant implements cosine as a dot product — it normalizes on upsert and compares with dot — so
|
| 87 |
+
`COSINE` costs the same as `DOT` here without assuming the caller preserved unit norm.
|
|
|
|
|
|
|
| 88 |
|
| 89 |
+
The table itself can also live in Qdrant, as a retrieve-by-id collection of one point per vocab
|
| 90 |
+
row (`hnsw_config=models.HnswConfigDiff(m=0)` — indexing it is pure waste), so the query path holds
|
| 91 |
+
no model weights at all.
|
|
|
|
|
|
|
|
|
|
| 92 |
|
| 93 |
+
### Without FastEmbed
|
|
|
|
|
|
|
| 94 |
|
| 95 |
+
`zero_encoder.py` is the reference implementation — 93 lines, `numpy` and `tokenizers`, no torch.
|
| 96 |
+
This downloads the whole repo, not just the 31 MB graph.
|
| 97 |
|
| 98 |
+
```python
|
| 99 |
+
from huggingface_hub import snapshot_download
|
| 100 |
+
import sys, numpy as np
|
| 101 |
+
|
| 102 |
+
d = snapshot_download("DylanCouzon/constella-zero")
|
| 103 |
+
sys.path.insert(0, d)
|
| 104 |
+
from zero_encoder import ZeroQueryEncoder
|
| 105 |
+
|
| 106 |
+
enc = ZeroQueryEncoder(d, variant="int8") # or "fp16"
|
| 107 |
+
q_np = enc.encode(["how do mrna vaccines work?"]) # (1, 1024), L2-normalized
|
| 108 |
+
assert np.abs(q_np[0] - q).max() < 1e-5 # the vector FastEmbed just produced
|
| 109 |
+
```
|
| 110 |
+
|
| 111 |
+
## How it works
|
| 112 |
+
|
| 113 |
+
Tokenize (WordPiece, special tokens on, truncate at 512, no padding, no prefix). A token appearing
|
| 114 |
+
`c` times carries **total weight `sqrt(c)`** — repetition saturates. Sum the rows, divide by the
|
| 115 |
+
weight sum, L2-normalize. An empty or near-zero-norm bag falls back to the normalized `[CLS]` row
|
| 116 |
+
(id 101). Per-token learned weights are folded into the rows, so the artifact is self-contained.
|
| 117 |
+
|
| 118 |
+
Because pooling is not a masked mean, it is done inside the ONNX graph rather than by the caller.
|
| 119 |
+
`config.json` carries the rule and its fingerprint (`adb24fb2e8cad66f`).
|
| 120 |
+
|
| 121 |
+
`int8` is the variant every number below was measured on; it is loss-free against `fp16` to within
|
| 122 |
+
0.00013 nDCG@10.
|
| 123 |
|
| 124 |
## Files
|
| 125 |
|
| 126 |
+
You need exactly one of these three.
|
| 127 |
+
|
| 128 |
+
| file | for | size |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 129 |
|---|---|---|
|
| 130 |
+
| `model.onnx` | FastEmbed, or any ONNX runtime — pooled and normalized, `(b, 1024)` | 31 MB |
|
| 131 |
+
| `model_tokens.onnx` | pipelines that insist on pooling themselves, `(b, s, 1024)` | 31 MB |
|
| 132 |
+
| `model.npz` | the numpy reference path | 94 MB |
|
|
|
|
|
|
|
| 133 |
|
| 134 |
+
Both graphs are opset 17, standard operators only, carrying the table as an int8 initializer with a
|
| 135 |
+
per-row fp32 scale dequantized in-graph.
|
|
|
|
|
|
|
| 136 |
|
| 137 |
+
The bundled tokenizer files are stella's, with `model_max_length`/`max_length` set to **512** and
|
| 138 |
+
`padding` to **null** — the rule the document index was built with. stella ships 32768/8000 and
|
| 139 |
+
fixed-512 padding, which any loader honouring those fields would otherwise apply.
|
| 140 |
+
`config.json` records the originals under `tokenizer_deviation_from_teacher`.
|
| 141 |
|
| 142 |
## Results
|
| 143 |
|
| 144 |
+
nDCG@10 on six BEIR datasets, exact search so ANN recall is not a confound. Measured once, on the
|
| 145 |
+
table shipped here (sha `a7007b1a…`).
|
|
|
|
| 146 |
|
| 147 |
+
| system | arguana | fiqa | nfcorpus | scidocs | scifact | trec-covid | **average** |
|
| 148 |
|---|---|---|---|---|---|---|---|
|
| 149 |
+
| **constella-zero (int8)** | 0.5916 | 0.3728 | 0.3124 | 0.1677 | 0.6101 | 0.5490 | **0.4339** |
|
| 150 |
+
| + BM25, convex fusion | 0.5975 | 0.4026 | 0.3497 | 0.1881 | 0.7068 | 0.7018 | **0.4911** |
|
| 151 |
| BM25 alone | 0.4878 | 0.2532 | 0.3180 | 0.1565 | 0.6791 | 0.6099 | 0.4174 |
|
| 152 |
+
| the teacher, used on both sides | 0.6369 | 0.5536 | 0.4134 | 0.2395 | 0.7796 | 0.8234 | 0.5744 |
|
| 153 |
|
| 154 |
+
A lookup table retains **75.5%** of the teacher's quality (0.4339 / 0.5744), with a query side
|
| 155 |
+
that does no matrix multiplication at all.
|
| 156 |
|
| 157 |
+
The fused row is **weighted score fusion**, `0.8 × dense + 0.2 × BM25` on min-max normalized
|
| 158 |
+
scores — not reciprocal rank fusion, so Qdrant's `Fusion.RRF` will not reproduce it.
|
|
|
|
|
|
|
|
|
|
| 159 |
|
| 160 |
+
## Limits
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 161 |
|
| 162 |
- **Teacher contamination.** stella discloses **ArguAna** and **FiQA** in its training data —
|
| 163 |
+
two of the six above, and ArguAna is its second-highest score. On the four sets with no
|
| 164 |
+
disclosed overlap it averages **0.4098 against BM25's 0.4409** — below BM25. Weight the
|
| 165 |
+
average accordingly.
|
| 166 |
+
- **It is a bag of tokens.** Word order, negation and syntax are not represented: "dog bites man"
|
| 167 |
+
and "man bites dog" give the same vector.
|
| 168 |
+
- **Out of domain it drops.** Training was Wikipedia- and e-commerce-shaped, and the six sets
|
| 169 |
+
above are further from that than the data it was fitted on.
|
| 170 |
+
- **English only**, 512 wordpieces, 30,522-token WordPiece vocab. Out-of-vocabulary terms degrade
|
| 171 |
+
to subword rows.
|
| 172 |
+
- **The document side is not cheap** — 2.05 GB per 1M documents at 1024-d fp16. The whole trade is
|
| 173 |
+
on the query side.
|
| 174 |
|
| 175 |
## Costs
|
| 176 |
|
| 177 |
| | |
|
| 178 |
|---|---|
|
| 179 |
+
| query asset (int8 rows + scales + tokenizer) | 31.8 MB |
|
| 180 |
+
| `model.onnx`, batch 1, one thread, 8-token query | 0.047 ms |
|
| 181 |
+
| `model.onnx`, batch 1, one thread, 512-token query | 1.22 ms |
|
| 182 |
+
| `zero_encoder.py`, batch 1, one CPU core | 0.38 ms |
|
| 183 |
+
| hydration (cold load to first query) | 0.22 s |
|
| 184 |
+
| document vectors, 1024-d fp16 / int8 | 2.05 / 1.02 GB per 1M — raw payload, before index overhead |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 185 |
|
| 186 |
+
The graph derives token counts from an all-pairs comparison, so cost grows with the **square** of
|
| 187 |
+
sequence length — 26x from an 8-token query to a 512-token one. Real queries sit at the short end
|
| 188 |
+
(median 13 wordpieces).
|
| 189 |
|
| 190 |
## Training
|
| 191 |
|
| 192 |
+
L2 regression of the table's pooled output onto the teacher's query embeddings, over 340,850
|
| 193 |
+
pairs plus 220,632 query-text-only rows, from **Amazon ESCI**, **FEVER**, **HotpotQA**, **SQuAD**,
|
| 194 |
+
**NQ-open**, **TriviaQA** and **Mr. TyDi (en)**. No MS MARCO.
|
| 195 |
|
| 196 |
+
Attribution: NQ, SQuAD, HotpotQA, FEVER and Mr. TyDi are Wikipedia-derived and **CC BY-SA**
|
| 197 |
+
(3.0/4.0); Amazon ESCI and TriviaQA are Apache-2.0; the teacher is MIT.
|
|
|
|
| 198 |
|
| 199 |
+
## Provenance
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 200 |
|
| 201 |
```
|
| 202 |
run_id p35w-2m-s2500
|
|
|
|
| 205 |
preproc prefix="" · add_special_tokens · max_length=512 · pool_mode=sqrt
|
| 206 |
preproc fingerprint adb24fb2e8cad66f
|
| 207 |
```
|
| 208 |
+
|
| 209 |
+
Published as `zero-query-encoder-v1` and renamed on 2026-09-03; the old URL redirects.
|