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

asmr-qwen3.5-9b-zh-tw-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.

Standard Mode

The traditional output format where only the translated text is returned.

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

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>, ...], "text": "<繁體中文>", "start": <最早ms>, "end": <最晚ms>}

字幕版權行：{"ids": [<n>], "text": null, "start": <ms>, "end": <ms>}
每個輸入id必須恰好出現在一個輸出項中。

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], "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-gguf-v0.2-q4_k_m.gguf -c 8192 --port 8080

llama-cli

llama-cli -m asmr-qwen3.5-9b-zh-tw-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-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
      },
      "text": {
        "anyOf": [
          {
            "type": "string"
          },
          {
            "type": "null"
          }
        ]
      },
      "start": {
        "type": "integer"
      },
      "end": {
        "type": "integer"
      }
    },
    "required": [
      "ids",
      "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: oproj, vproj, kproj, gateproj, qproj, upproj, down_proj
• —Locale: zh-tw (Traditional Chinese)
• —Mode: Standard Mode
• —Max sequence length: 8192
• —Precision: bf16

Version 0.2 — Native 8k Continuation

This release continues the v0.1 Traditional Chinese standard 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, 120 far-context/glossary-recall, and 160 ordinary replay draws. The published adapter reached absolute step 200. Adapter weights SHA-256: d408a36cfd8e5eaf2690c5668733ea61c5f63c55dbd1bc0c38d374f8319dffe9.

Pinned base revision: 005429cee5cb648998cf2b70eebdd83175989c9a.

Validation Scope

This variant received a small held-out semantic check rather than the full 200-row matrix: three ordinary rows and three long rows, each from a distinct locked-test family and with zero source overlap with the continuation ledger. All six grammar-constrained Q8 requests completed with exact ID coverage.

Manual reading found five of six rows usable or usable with local correction. One severely corrupted/repetitive long-ASR row produced finite generic repetition and heavy omission. Re-evaluating its three exact original 4k constituent prompts reduced the repetition substantially, but did not eliminate invented fallback text on unusable ASR.

Context Budgeting

8192 is the total runtime context, not a safe input-only budget. Reserve room for generated JSON. For noisy or repetitive ASR, shorter windows are more robust than concatenating material to the context limit.

Known Limitations

• —This variant did not receive the full 100-row 4k plus 100-row 8k validation used for zh-tw-echo.
• —Corrupted long ASR can trigger finite repetition, omission, and invented scene continuation.
• —The small sample showed entity-name confusion in one long row.
• —There is no complete controlled v0.1-versus-v0.2 final comparison.
• —Validation used Q8_0; 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.