mmis1000/asmr-qwen3.5-9b-zh-tw-echo-gguf-v0.2

asmr-qwen3.5-9b-zh-tw-echo-gguf-v0.2

GGUF quantizations of a fine-tuned model for translating Japanese ASMR transcriptions (ASR/Whisper output) into Traditional Chinese.

The model normalizes imperfect audio transcriptions, applies domain-specific glossaries, and translates character dialogue while retaining emotion and nuances.

This variant preserves the upstream Qwen3.5 MTP / speculative-decoding head in GGUF format so it can be used with MTP-capable llama.cpp builds.

Echo Mode

The model echoes the source Japanese text in an "input" field and records applied terms in a per-entry "glossary" object alongside the target translation. This provides an explicit source anchor that can reduce omitted or drifted segments, but it does not guarantee immunity to long-context repetition or noisy-ASR failures.

Available Quantizations

Prompt Example

將以下日語ASMR逐字稿翻譯成繁體中文。

音軌：track01_示例音軌
場景說明：主角與青梅竹馬在校園下午的對話...

術語表（請嚴格使用zh欄位的譯名）：
{
  "cvs": [],
  "characters": [],
  "terms": [{"ja": "放課後", "zh": "放學後"}]
}

翻譯前請靜默修正下列Whisper識別錯誤：
- 重複片語（連續3次以上且無變化）：僅保留一次
- 錯字／同音異字：依上下文修正
- 字幕版權行（字幕：／翻訳：／QQ／LINE水印）：text設為null
- 錯誤專有名詞：依術語表修正

翻譯規則：
- 呻吟與氣息聲（あ、ん、はあ）→ 自然對應（啊、嗯、哈、呼）
- 擬聲詞：日語形式翻譯（パンパン→啪啪）；中文形式保留原樣
- 保留角色語氣與口吻
- text欄位只輸出譯文，不加注釋或括號說明
- input欄位為ids所對應的原始日文片段
- glossary欄位只記錄該輸出項實際使用的日文→中文術語；沒有時輸出空物件

輸入：逐字稿JSON陣列 — {"id": <n>, "text": "<日文>", "start": <ms>, "end": <ms>}

輸出：將連續構成同一句話的片段合併，JSON陣列格式：
{"ids": [<n>, ...], "input": "<合併後的原始日文，以空格連接>", "glossary": {"<日文>": "<繁體中文>"}, "text": "<繁體中文>", "start": <最早ms>, "end": <最晚ms>}

字幕版權行：{"ids": [<n>], "input": "<原始日文>", "glossary": {}, "text": null, "start": <ms>, "end": <ms>}
每個輸入id必須恰好出現在一個輸出項中。
input欄位為ids所對應的原始日文片段以空格連接，glossary記錄該項實際使用的術語，text為其繁體中文翻譯。

逐字稿：
[
  {"id": 1, "text": "ねぇ、放課後、", "start": 3000, "end": 5000},
  {"id": 2, "text": "一緒に帰らない?", "start": 5000, "end": 7000}
]

Example Output:

[{"ids": [1, 2], "input": "ねぇ、放課後、 一緒に帰らない?", "glossary": {"放課後": "放學後"}, "text": "欸，放學後，要不要一起回去？", "start": 3000, "end": 7000}]

Usage

llama-server

llama-server -m asmr-qwen3.5-9b-zh-tw-echo-gguf-v0.2-q4_k_m.gguf -c 8192 --port 8080

llama-cli

llama-cli -m asmr-qwen3.5-9b-zh-tw-echo-gguf-v0.2-q4_k_m.gguf -c 8192 -p "<your prompt>" -n 2048

llama-cli with MTP

llama-cli -m asmr-qwen3.5-9b-zh-tw-echo-gguf-v0.2-q4_k_m.gguf -c 8192 --spec-type draft-mtp --spec-draft-n-max 6 -p "<your prompt>" -n 2048

Structured Decoding (Recommended)

This model outputs JSON arrays. Using structured decoding (e.g. GBNF grammar or JSON schema constraints) avoids wasted computation on malformed output and guarantees valid JSON on every generation.

JSON Schema:

{
  "type": "array",
  "items": {
    "type": "object",
    "properties": {
      "ids": {
        "type": "array",
        "items": {
          "type": "integer"
        },
        "minItems": 1
      },
      "input": {
        "type": "string"
      },
      "glossary": {
        "type": "object",
        "additionalProperties": {
          "type": "string"
        }
      },
      "text": {
        "anyOf": [
          {
            "type": "string"
          },
          {
            "type": "null"
          }
        ]
      },
      "start": {
        "type": "integer"
      },
      "end": {
        "type": "integer"
      }
    },
    "required": [
      "ids",
      "input",
      "glossary",
      "text",
      "start",
      "end"
    ],
    "additionalProperties": false
  },
  "minItems": 1
}

Supported by llama.cpp (--json-schema), vLLM, and outlines.

Training Details

• —Base model: unsloth/Qwen3.5-9B
• —Method: LoRA (r=16, alpha=16)
• —Target modules: gateproj, qproj, vproj, upproj, downproj, oproj, k_proj
• —Locale: zh-tw (Traditional Chinese)
• —Mode: Echo Mode
• —Max sequence length: 8192
• —Precision: bf16

Version 0.2 — Native 8k Continuation

This release continues the v0.1 Traditional Chinese echo adapter directly to a maximum training sequence length of 8,192 tokens. It does not stack adapters and does not use RoPE scaling or YaRN.

The deterministic continuation consumed 800 training draws:

• —520 coherent same-track long-context draws;
• —120 far-context/glossary-recall draws;
• —160 ordinary replay draws.

The published adapter reached absolute step 200. Adapter weights SHA-256: c6f79936c2cf65bea626f61c350d6318b1578b7ef7d2860d0d8f4619cc0ff5df.

Pinned base revision: 005429cee5cb648998cf2b70eebdd83175989c9a.

Validation

This is the only v0.2 variant that received the full held-out matrix: 100 ordinary-context rows and 100 long-context rows, with exact stored prompts and grammar-constrained Q8 inference.

Position-quartile chrF for completed 8k output was 0.2558 / 0.2695 / 0.2575 / 0.2545; no late-context quality cliff was observed.

Nine direct 8k requests exhausted the 8,192-token runtime context before completing JSON. Those fixtures fit as training sequences, but generated translations can be longer than their references. The test harness deliberately did not use application-level retry or sub-chunk recovery.

Context Budgeting

8192 is the total runtime context, not a safe input-only budget. Reserve room for the generated JSON. Do not fill the entire context with the prompt and expect a non-trivial response; split the transcript or use an application retry/sub-chunk strategy.

Known Limitations

• —Grammar-constrained JSON verifies structure, not translation quality.
• —Long context increased false-null behavior on short fillers/interjections in the held-out matrix.
• —Corrupted or repetitive ASR can still cause condensation, invention, or repetition.
• —The v0.2 final checkpoint was not run in a complete controlled v0.1-versus-v0.2 comparison.
• —Validation used the Q8_0 export; other quantizations were not independently quality-scored.

Content Notice

The training domain includes adult ASMR dialogue and may produce sexually explicit text. This model is intended for transcription translation and subtitle-processing workflows.

MTP-preserving export

This revision preserves and verifies all 15 upstream MTP tensors against the pinned base checkpoint after the LoRA merge. The native-8192 adapter is unchanged; no retraining was performed. The MTP head itself is not fine-tuned. Earlier revisions of this v0.2 repository omitted MTP. Filenames and translation weights remain compatible; use a fresh repository revision to avoid cached non-MTP files. All four quantizations have verified MTP metadata and tensor inventory. Q4KM was GPU-smoke-tested with llama.cpp b9247, context 8192, and --spec-type draft-mtp --spec-draft-n-max 3. This is a runtime smoke test, not a new quality evaluation or a speedup guarantee. See mtp-preservation.json and mtp-release-verification.json for provenance and checks.