Models

Winnow can run two kinds of models. Both come from the Squeez project's approach and are trained on the same data: tool outputs from SWE-bench agent runs, labelled with the lines the agent actually needed.

	pooled (default)	highlighter
Model	32M line classifier, fine-tuned from jhu-clsp/ettin-encoder-32m	KRLabsOrg/verbatim-rag-modern-bert-v2, 150M span model
Output	a probability per line	character spans
Where to get it	rbk4209/winnow-pooled-32m on the Hub (128 MB)	downloaded from the Hub on first use
Median latency¹	0.29 s	1.5 s
p95 latency¹	0.92 s	9.0 s
Recall as shipped¹	0.847	0.836
Token reduction¹	41.4%	39.7%

¹ Full Squeez test split, consumer GPU, no token budget.

Pooled

The pooled model reads the task and the tool output in one pass, with a separator token between lines, and mean-pools each line's tokens into a relevance score. It was trained for this project with training/kaggle_train_pooled.ipynb and matches the highlighter's recall at a fraction of its cost, which lets Winnow score practically every tool output on a GPU instead of only small ones. It is the default: nothing needs to be configured. To use your own pooled model:

export HEADROOM_WINNOW_MODEL=/path/to/your/pooled-model

Highlighter

Squeez's original extractive model. Select it with:

export HEADROOM_WINNOW_BACKEND=highlighter

Loading is pinned to a reviewed commit (weights and tokenizer) because the model runs its own code via trust_remote_code. Its latency grows quickly with input length, so its default budget is 2,048 tokens on a GPU.

Bring your own

Any object with a find_spans(query, content) -> list[tuple[int, int]] method works as a backend:

from headroom_winnow import WinnowCompressor


class KeywordBackend:
    def find_spans(self, query: str, content: str) -> list[tuple[int, int]]:
        spans, start = [], content.find("ERROR")
        while start != -1:
            spans.append((start, start + 5))
            start = content.find("ERROR", start + 1)
        return spans


winnow = WinnowCompressor(KeywordBackend())

Give it a max_input_tokens attribute to set a token budget.