---
license: apache-2.0
language:
  - zh
  - en
pretty_name: ScholarSearchAgent Auxiliary Models
tags:
  - information-retrieval
  - reranking
  - text-classification
  - lora
  - academic-search
---

# ScholarSearchAgent Auxiliary Models

This dataset repository hosts the trained auxiliary model artifacts used by
ScholarSearchAgent, an intelligent academic paper search and recommendation
system for complex research queries.

Source code: <https://github.com/Kaedeser/ScholarSearchAgent>

## Download And Restore

The complete model bundle is stored as 200 MiB archive parts named
`ScholarSearchAgentAuxiliaryModel_20260723.tar.part0001`,
`ScholarSearchAgentAuxiliaryModel_20260723.tar.part0002`, and so on. Download
all parts in numerical order, concatenate them, verify the SHA-256 value, then
extract the resulting TAR archive.

PowerShell:

```powershell
Get-ChildItem "ScholarSearchAgentAuxiliaryModel_20260723.tar.part*" |
  Sort-Object Name |
  Get-Content -AsByteStream |
  Set-Content -AsByteStream "ScholarSearchAgentAuxiliaryModel_20260723.tar"

Get-FileHash "ScholarSearchAgentAuxiliaryModel_20260723.tar" -Algorithm SHA256
tar -xf "ScholarSearchAgentAuxiliaryModel_20260723.tar"
```

Expected SHA-256:

```text
c15830db9b3c126d584b1ffbdef17cd70a4101d20470ad78661eb883b4fe4a46
```

## Included Models

| Directory after extraction | Purpose | Model form | Reported result |
| --- | --- | --- |
| `models/query_gate_biobert` | Detect whether a query should enter the academic-search pipeline | BioBERT sequence classifier | Test accuracy: 0.997143 |
| `models/intent_biobert` | Classify academic search intent | BioBERT sequence classifier | See packaged training metadata |
| `models/selector_reranker` | Rerank recalled paper candidates | Sentence Transformers CrossEncoder | F1: 88.81% |
| `models/crawler_strategy_lora` | Select section-expansion strategies | Qwen2.5-3B-Instruct LoRA adapter | Section F1: 0.3007 |

The crawler strategy artifact is a LoRA adapter. The public
`Qwen2.5-3B-Instruct` base model is not redistributed in this repository and
must be obtained separately under its original license.

## Data And Privacy

The archive contains model weights, tokenizer/configuration files, and training
metadata required for inference. It does not contain API keys, passwords,
private endpoints, raw training datasets, source-code history, or participant
identity information.