About

This model was converted to GGUF format from Surpem/Supertron2-Reranker-8B using a modified version of llama.cpp (release b10269). Refer to the original model card for more details on the model.

This is a working GGUF. It will work on any modern version of llama.cpp.

Most community GGUFs of Supertron2-Reranker-8B produce garbage scores because of 4 issues:

1. Missing classification head. (This is 99% of the problem)
2. Wrong Token Extracted -> yes (9693) and no (2152) must come from the output head, which is then dropped. This GGUF is rerank-only.
3. Wrong Scoring Formula -> needs sigmoid(yes_logit - no_logit).
4. Different chat template.

This GGUF fixes all 4 problems.

Original Model Card

Supertron2-Reranker-8B: A Compact Cross-Encoder Reranking Model

Model Description

Supertron2-Reranker-8B is a reranking model built on top of Qwen/Qwen3-VL-Reranker-8B. It is designed to score query-document pairs for retrieval pipelines, search systems, and RAG applications where a stronger second-stage ranker is useful.

• Developed by: Surpem
• Model type: Cross-Encoder Reranker
• Architecture: Qwen3-VL reranker, 8B parameters
• License: Apache 2.0

---

Capabilities

Search Reranking

Supertron2-Reranker-8B can compare a user query against candidate passages and assign relevance scores. It is intended as a second-stage reranker after a faster retriever has already selected candidate documents.

RAG Pipelines

The model can help improve retrieval-augmented generation by pushing more relevant documents toward the top of the context window before answer generation.

Question-Document Matching

Supertron2-Reranker-8B is useful for matching questions to passages, snippets, help-center articles, documentation chunks, and other text candidates.

Instruction-Aware Retrieval

The model is prompted for relevance scoring, making it suitable for natural language search tasks where query intent matters.

---

Get Started

from sentence_transformers import CrossEncoder

model_id = "Surpem/Supertron2-Reranker-8B"

model = CrossEncoder(model_id)

pairs = [
    ("What is the capital of France?", "Paris is the capital and largest city of France."),
    ("What is the capital of France?", "Mars is often called the red planet."),
]

scores = model.predict(pairs)
print(scores)

Example reranking:

query = "How do I reset my password?"
documents = [
    "Use the account recovery page to reset your password.",
    "Our refund policy allows returns within 30 days.",
    "Two-factor authentication adds extra login security.",
]

results = model.rank(query, documents)
print(results)

---

Hardware Requirements

For larger batches or long documents, use more VRAM or reduce the batch size/max sequence length.

---

Intended Use

Supertron2-Reranker-8B is intended for:

• Search reranking
• RAG document reranking
• Query-passage relevance scoring
• Documentation and knowledge-base retrieval
• Evaluation of candidate retrieval results

It is not intended to be used as a standalone chat model.

---

Limitations

• The model scores relevance; it does not generate answers.
• It should be evaluated on your own retrieval domain before production use.
• Long documents may need chunking before reranking.
• Relevance scores are relative and may not be calibrated across unrelated queries.
• The model may still rank incorrect, outdated, or unsafe content highly if it appears textually relevant.

---

Citation

@misc{surpem2026supertron2-reranker-8b,
      title={Supertron2-Reranker-8B -- Compact Cross-Encoder Reranking Model},
      author={Surpem},
      year={2026},
      url={https://huggingface.co/Surpem/Supertron2-Reranker-8B},
}