Voxtral Mini 4B Realtime 2602 - modified decoder graph

This directory contains a modified version of the file
  onnx/decoder_model_merged_q4f16.onnx
from the repository onnx-community/Voxtral-Mini-4B-Realtime-2602-ONNX
(revision db34633130c460bd650644fc18dc90710d9f3e38), an ONNX export of
Voxtral Mini 4B Realtime 2602 by Mistral AI (mistralai/Voxtral-Mini-4B-Realtime-2602).
Both are licensed under the Apache License, Version 2.0 (see LICENSE.txt in this directory).

Modified file: decoder_model_merged_q4f16_d2400.onnx
SHA-256: 3303c959ce35a43a765d48f326539cc0a75b6e8ffff7c46678586cf1317ea5f5

Changes (made by the operator of transcriber.tools.ryusei.io on 2026-10-28):
- The 26 constants /model/layers.N/ada_scale (FLOAT16 [1, 3072]), which condition the model
  on its transcription delay, were recomputed for a delay of 30 tokens (2.4 seconds) instead of
  the 6 tokens (480 ms) the original file was exported with. They were computed from the
  delay-conditioning weights of the original model, and the new values are stored in this file.
- All other initializers and all nodes are unchanged. The external data file
  (decoder_model_merged_q4f16.onnx_data) is used unmodified from the original repository.

----

このディレクトリの decoder_model_merged_q4f16_d2400.onnx は、上記の原本（Apache License 2.0。全文は LICENSE.txt）を改変したものです。
文字起こしの遅延を表す 26 個の定数（/model/layers.N/ada_scale）を、配布時の 6 トークン（480 ms）から
30 トークン（2.4 秒）向けに原本のモデルの重みから計算し直し、このファイルの中に収めました。
ほかの定数とノードは原本と同じで、外部データ（decoder_model_merged_q4f16.onnx_data）は原本のものを改変せずに使います。
改変した者と日付: transcriber.tools.ryusei.io の運営者、2026-10-28