Murasaki-Project/Murasaki-APE-Aligner-2B

VISIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2.3BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Mar 9, 2026License:mitArchitecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Murasaki-Project/Murasaki-APE-Aligner-2B is a 2.3 billion parameter language model based on Qwen-3.5-2B-Base, specifically designed for post-processing and alignment tasks in ACGN text translation. It excels at repairing control codes and formatting in translated game scripts, handling over 1000 types of game control characters from more than 40 engines. This model focuses solely on correcting code and format, ensuring logical and syntactically correct output without altering translation content. It is optimized for environments with sufficient VRAM or as a base for further fine-tuning.

Loading preview...

Murasaki-APE-Aligner-2B: Specialized Format Repair for ACGN Translations

Murasaki-APE-Aligner-2B is a 2.3 billion parameter model, built upon Qwen-3.5-2B-Base, engineered for the unique challenges of ACGN (Anime, Comics, Games, Novels) text translation. Its primary function is to post-process and align translated game scripts, particularly those containing complex control codes found in Galgame and RPG engines.

Key Capabilities

  • Extensive Control Code Handling: Trained on data from over 40 Galgame and RPG engines, encompassing more than 1000 distinct control character formats.
  • Automated Format Repair: Capable of correcting misaligned, missing, redundant, or incorrectly translated control parameters.
  • Preserves Translation Content: Crucially, the model only modifies code and formatting, ensuring the integrity of the translated text's meaning remains untouched.
  • JSON Input/Pure Text Output: Processes JSON inputs containing original (ja) and draft translated (zh) text, outputting a clean, repaired plain text string.
  • High Precision: The provided model weights are in BF16 (4.45 GB), suitable for environments with ample VRAM or as a foundation for further fine-tuning.

When to Use This Model

This model is ideal for developers and translators working with:

  • Game Localization: Specifically for Galgame, RPG, and similar interactive media where text is interleaved with complex formatting and control codes.
  • Automated Translation Pipelines: To integrate a robust post-processing step that ensures the technical correctness of translated game scripts.
  • Fixing Common Translation Errors: Addresses issues like [font color=0x64E560]respect[resetfont][resetfont] where a [resetfont] might be duplicated or misplaced.

For optimal stability, it is recommended to use low temperature values (0.0-0.1) during inference.

Finetunes
3 models