Chopper 0.31B: reranker and context compressor in one pass

0.31B parameters · about 50 ms per question on an RTX 3090 (20 snippets, ~2,400 tokens) · MIT

Reads the passages your retriever found and, in a single forward pass, (1) scores every passage, like a reranker, and (2) cuts out chops: the exact sentences that answer the question, typically 2–5% of the text, never rewritten.

This repository holds the same weights as the mmbert-base/ folder of wiiiktor/chopper_mmbert (the main repository: all variants, full results), packaged on its own so leaderboards can list it by size.

Also a reranker, in the same pass. Every passage gets a relevance score (mean p(KEEP) of its 16 most likely tokens):

nDCG@10, 8-language rerank holdout (400 queries)
this model 0.973
bge-reranker-v2-m3 0.984
Qwen3-Reranker-0.6B 0.977
# pip install chopper-rag
from chopper_rag import Chopper
ch = Chopper.from_pretrained(repo="wiiiktor/chopper-0.31B", variant="base")
question = "Who founded the company that makes the drink?"
out = ch.run(question, passages)   # passages: list of strings from your retriever
out.chops                        # exact quotes
ch.rerank(question, passages)    # [(passage_index, score), ...], best first

Code, training recipe and benchmarks: github.com/wiiiktor/chopper_demo.

Downloads last month
19
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for wiiiktor/chopper-0.31B

Finetuned
(210)
this model

Dataset used to train wiiiktor/chopper-0.31B