---
license: cc-by-nc-4.0
license_link: LICENSE
base_model: Qwen/Qwen3-0.6B-Base
library_name: transformers
tags:
  - generative-retrieval
  - semantic-id
  - navigational-search
  - web-search
---

# Octen-WebGR-0.6B

[![Discord](https://img.shields.io/badge/Discord-Join%20Octen-5865F2?logo=discord&logoColor=white)](https://discord.gg/Y3Yq5qxk7h)

Octen-WebGR-0.6B is a small generative retriever for navigational web search, developed by [Octen](https://octen.ai/). Given a query, it generates a two-token semantic
identifier (SID) that is constrained to a table of 30.9 million real URLs, so the web index lives in the model's parameters. It needs
1.6 GiB of GPU memory and also runs on CPUs (72 ms per query on one server socket).

It is described in the technical report [*Octen-WebGR: Scaling Generative Retrieval to 200 Million Web Pages*](https://arxiv.org/abs/2610.35256) (Xu et al., 2026).
The larger model for a 196-million-URL table is [Octen-WebGR-4B](https://huggingface.co/Octen/Octen-WebGR-4B). The 31M universe of
this model is a subset of the 4B model's universe and uses the same identifiers.

## Model

| | |
|---|---|
| Base model | Qwen3-0.6B-Base, full-parameter fine-tuning |
| Layers / hidden size | 28 / 1,024 |
| Parameters, total (non-embedding) | 0.73B (0.44B) |
| Identifier tokens | 2 × 65,536 (vocabulary ids 151,936 to 283,007) |
| URLs in the training universe | 31,282,165 |
| URLs in the released table | 30,931,888 (after screening, see below) |
| Training data | Query–URL click pairs up to August 2026 (proprietary, not released) |
| Weights | bf16, 1.5 GB |

## Results

Navigational queries from September 2026, the month after the training data, weighted by clicks. A system scores a hit when the
clicked target page is among its top *k* URLs; targets outside a system's candidate set count as misses, so this model, which only
returns URLs from its 31M universe, is scored against the same targets as the systems over the 201M universe.

| System | hit@1 | hit@10 |
|---|---|---|
| Octen-WebGR-4B (201M-URL universe) | **59.6** | **68.7** |
| **Octen-WebGR-0.6B** | 56.5 | 64.0 |
| Dense retrieval (Octen embedding + adapter trained on the same clicks, exact search, 201M) | 24.8 | 55.6 |
| Dense retrieval + click prior | 46.5 | 64.2 |
| BM25 over URL texts (URL + training queries) | 9.1 | 25.3 |
| Lookup in the training click table | 58.9 | 62.5 |

The 0.6B model is within 3.1 points of the 4B model at hit@1 and matches it on new queries for familiar pages, but it finds few
rarely visited pages, most of which are outside its universe. All numbers use the full 31M universe; restricting decoding to the
screened table lowers hit@1 from 56.5 to 55.7. See the report for the full evaluation.

## How it works

1. Normalise the query: collapse whitespace, strip, lowercase. Keep at most the first 32 tokens.
2. Prompt: `Query: {q}\nURL:` without special tokens.
3. Generate two SID tokens. Level *t* (0 or 1) uses vocabulary id `151936 + t * 65536 + code`.
4. Constrained beam search: at each level, only prefixes present in `sid_table/` are allowed and the softmax is taken over the valid
   children only; 20 beams are kept. The score is the sum of the two log-probabilities.
5. Look `(sid0, sid1)` up in `sid_table/` to get the URL.

> **Do not use `model.generate()`.** The tokenizer has no entries for the SID tokens, and unconstrained generation returns
> identifiers that map to no URL. Use the inference code in this repository.

## Usage

```bash
pip install -r requirements.txt      # torch 2.8.0, transformers 4.57.1; flash-attn optional
python infer_cpu.py --model Octen/Octen-WebGR-0.6B --query "youtube" "boeing internship" --threads 16
python infer_gpu.py --model Octen/Octen-WebGR-0.6B --query "youtube"
python infer_cpu.py --model Octen/Octen-WebGR-0.6B --check --threads 16    # compare against smoke_test.json
```

```python
from huggingface_hub import snapshot_download
import sys

root = snapshot_download("Octen/Octen-WebGR-0.6B")
sys.path.insert(0, root)
from infer_cpu import GRCPU

gr = GRCPU(root)                     # loads the model and the 30.9M-row table
for url, logp, row in gr.search(["hugging face"], K=10)[0]:
    print(url, logp)
```

`infer_cpu.py` reuses the prompt's key–value cache for the second level and takes the softmax over valid children only. On one
socket of an Intel Xeon 8568Y+ (bf16, cores pinned) it answers a single query in 72 ms (median) and about 57 queries per second in
batches of 8; its hit@1 is within 0.1 points of the GPU implementation. `infer_gpu.py` is a reference implementation that recomputes
the prompt for every beam; on an idle H200 it answers a single query in 43.5 ms and needs 3.9 ms per query in batches of 32.

Memory: 1.6 GiB of GPU memory. Peak host RAM is about 4.5 GiB with `infer_cpu.py`, which keeps the weights and the table in memory.

## Files

| Path | Content | License |
|---|---|---|
| `model.safetensors`, `config.json`, `generation_config.json` | Model weights (bf16) | CC BY-NC 4.0 |
| `tokenizer*.json`, `vocab.json`, `merges.txt` | Qwen3 tokenizer, unchanged | Apache 2.0 |
| `sid_table/part-*.parquet` | One row per URL: `sid0`, `sid1` (int32), `url` (string, without scheme), sorted by `sid0 * 65536 + sid1`, unique | CC BY-NC 4.0 |
| `infer_gpu.py`, `infer_cpu.py` | Inference code | Apache 2.0 |
| `smoke_test.json` | 31 hand-picked queries with expected top-20 URLs, for `--check` | – |

## Screening of the table

Before release we removed every URL whose host, or a parent domain of it, is on the adult, malware, phishing, cryptojacking or
stalkerware lists of the UT1 blacklists; every URL listed by URLhaus; and every URL whose text contains an e-mail address or a
credential (350,277 URLs, 1.1%). Removing a row from the table removes the URL from every possible output, because decoding is
constrained to the table. The tables were not checked against lists of child sexual abuse material, nor screened for personal pages;
both are left to the takedown process.

URL removal requests: research@octen.ai. Removals are published as versioned changes to `sid_table/`.

## Intended use and limitations

- Intended for research on generative retrieval and for navigational search over the URLs in the released table.
- The model can only return URLs in the table. Pages that appeared after August 2026 are not covered, and adding a page requires
  training on it.
- It is weaker than dense retrieval on rarely visited pages, and it is not meant for informational queries.
- Its knowledge reflects clicks up to August 2026 and the biases of the rankings under which those clicks were made.
- It was trained on aggregated click data without user identifiers. It must not be used to infer anything about individual users.

## License

The model weights and the SID-to-URL table are released under [CC BY-NC 4.0](LICENSE) (non-commercial use). The inference code is
released under the [Apache License 2.0](LICENSE-CODE). The model is derived from Qwen3-0.6B-Base, Copyright 2024 Alibaba Cloud,
licensed under the Apache License 2.0 ([LICENSE-QWEN3](LICENSE-QWEN3)); see [NOTICE](NOTICE). Unlike the Octen-Embedding models, the
WebGR weights are not licensed for commercial use; contact Octen for commercial licensing.

## Citation

```bibtex
@misc{octen2026webgr,
  title         = {Octen-WebGR: Scaling Generative Retrieval to 200 Million Web Pages},
  author        = {Xu, Guangwei and Zhang, Wenkai and Liu, Yawen and Zhang, Dongxu and Xu, Xijie and Zou, Kuan},
  year          = {2026},
  eprint        = {2610.35256},
  archivePrefix = {arXiv},
  primaryClass  = {cs.IR},
  url           = {https://arxiv.org/abs/2610.35256}
}
```

## Contact

For API access, deployment solutions, and technical documentation, visit [octen.ai](https://octen.ai/).

Join our [Discord community](https://discord.gg/Y3Yq5qxk7h) for questions, feedback, and the latest updates.