Loading page...Загружаем страницу...正在加载页面...
AtlasRepo

RWKV-7 + KAN Adapter for Faithful Prose Rewriting

Description and use

RWKV‑7 with a KAN adapter rewrites dense prose while preserving numbers, identifiers, units, and negation, achieving perfect recall in diagnostic tests.

Problem

Rewriting dense prose can improve readability but often alters critical details such as numbers, identifiers, units, or negation, leading to correctness failures.

Outcome

The adapter achieves perfect recall on numbers, identifiers, negation, and units in the eight‑case diagnostic, with a 4‑gram copy score of 0.462, indicating substantial rewriting while preserving factual content.

Workflow

  • 1. Install torch and transformers.
  • 2. Load the frozen RWKV‑7 World3 1.5B base model.
  • 3. Attach the KAN adapter (kan_rwkv7_v7.2_best_adapter.pt).
  • 4. Use the provided prompt format and sampling settings to generate rewritten text.
  • 5. Review output for any factual discrepancies before deployment.

Evidence

The recorded eight‑case generation gate reported: Number recall 1.00, Identifier recall 1.00, Negation preservation 1.00, Completed generations 1.00, 4‑gram copy score 0.462.

Source

The adapter is 78.14 MB, and the repository pins the upstream base‑model revision used by the quick start.

Source

This project is for readable, faithful rewriting and research. It should not be used to misrepresent authorship, fabricate evidence, alter safety‑critical instructions, or conceal material changes.

Source

  • Record date: 2026-09-15T08:00:49.926Z