arjunkshah21/sc-keep-crossencoder-v4-large

SuperCompress Neural Keep (v4-large)

Open weights for SuperCompress v2 query-aware context compression (~395.83M parameters).

This cross-encoder scores context lines against the current query and keeps answer-critical evidence in original wording (selection, not summarization).

• Product: supercompress.dev
• Launch: engine-v2-launch
• Base model: answerdotai/ModernBERT-large
• Checkpoint id: sc-keep-crossencoder-v4-large
• Weights: model.safetensors (~1.58 GB)
• License: MIT

Intended use

Hosted / self-hosted Neural Keep for coding-agent and RAG context dumps: compress bulky tool output before it hits an LLM, while preserving required evidence.

Coding-agent benchmark (B5): 64.1% mean cut, required evidence in 24/24 cases.

Files

File	Role
model.safetensors	Trained keep / drop scorer
config.json	Model config
tokenizer.json / tokenizer_config.json	Tokenizer
sc_meta.json	SuperCompress training / gate metadata

Load

from transformers import AutoModelForSequenceClassification, AutoTokenizer

repo = "arjunkshah21/sc-keep-crossencoder-v4-large"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForSequenceClassification.from_pretrained(repo)

Production inference also runs via the SuperCompress API / MCP (https://www.supercompress.dev/api/mcp) without downloading weights.

Citation

SuperCompress v2 — open-weight Neural Keep. https://www.supercompress.dev/