---
base_model:
- google-bert/bert-base-uncased
datasets:
- microsoft/ms_marco
language:
- en
library_name: transformers
pipeline_tag: feature-extraction
license: mit
---

# Model Card
This is the official model from the paper [Hypencoder: Hypernetworks for Information Retrieval](https://arxiv.org/abs/2502.05364).

## Model Details
This is a Hypencoder Dual Encoder. It contains two trunks the text encoder and Hypencoder. The text encoder converts items into 768 dimension vectors while the Hypencoder converts text into a small neural network which takes the 768 dimension vector from the text encoder as input. This small network is then used to output a relevance score. To use this model please take a look at the [Github](https://github.com/jfkback/hypencoder-paper) page which contains the required code and details on how to run the model.

### Model Variants
We released the four models used in the paper. Each model is identical except the small neural networks, which we refer to as q-nets, have different numbers of hidden layers.

| Huggingface Repo | Number of Layers |
|:------------------:|:------------------:|
| [jfkback/hypencoder.2_layer](https://huggingface.co/jfkback/hypencoder.2_layer) |          2        |
| [jfkback/hypencoder.4_layer](https://huggingface.co/jfkback/hypencoder.4_layer) |          4        |
| [jfkback/hypencoder.6_layer](https://huggingface.co/jfkback/hypencoder.6_layer) |          6        |
| [jfkback/hypencoder.8_layer](https://huggingface.co/jfkback/hypencoder.8_layer) |          8        |

## Usage

```python
from hypencoder_cb.modeling.hypencoder import Hypencoder, HypencoderDualEncoder, TextEncoder
from transformers import AutoTokenizer

dual_encoder = HypencoderDualEncoder.from_pretrained("jfkback/hypencoder.6_layer")
tokenizer = AutoTokenizer.from_pretrained("jfkback/hypencoder.6_layer")

query_encoder: Hypencoder = dual_encoder.query_encoder
passage_encoder: TextEncoder = dual_encoder.passage_encoder

queries = [
    "how many states are there in india",
    "when do concussion symptoms appear",
]

passages = [
    "India has 28 states and 8 union territories.",
    "Concussion symptoms can appear immediately or up to 72 hours after the injury.",
]

query_inputs = tokenizer(queries, return_tensors="pt", padding=True, truncation=True)
passage_inputs = tokenizer(passages, return_tensors="pt", padding=True, truncation=True)

q_nets = query_encoder(input_ids=query_inputs["input_ids"], attention_mask=query_inputs["attention_mask"]).representation
passage_embeddings = passage_encoder(input_ids=passage_inputs["input_ids"], attention_mask=passage_inputs["attention_mask"]).representation

# The passage_embeddings has shape (2, 768), but the q_nets expect the shape
# (num_queries, num_items_per_query, input_hidden_size) so we need to reshape
# the passage_embeddings.

# In the simple case where each q_net only takes one passage, we can just
# reshape the passage_embeddings to (num_queries, 1, input_hidden_size).
passage_embeddings_single = passage_embeddings.unsqueeze(1)
scores = q_nets(passage_embeddings_single)  # Shape (2, 1, 1)
# [
#    [[-12.1192]],
#    [[-13.5832]]
# ]

# In the case where each q_net takes both passages we can reshape the
# passage_embeddings to (num_queries, 2, input_hidden_size).
passage_embeddings_double = passage_embeddings.repeat(2, 1).reshape(2, 2, -1)
scores = q_nets(passage_embeddings_double)  # Shape (2, 2, 1)
# [
#    [[-12.1192], [-32.7046]],
#    [[-34.0934], [-13.5832]]
# ]
```

## Citation
**BibTeX:**
```
@misc{killingback2025hypencoderhypernetworksinformationretrieval,
      title={Hypencoder: Hypernetworks for Information Retrieval}, 
      author={Julian Killingback and Hansi Zeng and Hamed Zamani},
      year={2025},
      eprint={2502.05364},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2502.05364}, 
}
```