Downloads · 30 days
708
60% of all-time downloads
the-walking-fish/jijun-LM-GGUF
jijun-LM-GGUF is a text generation model from the-walking-fish. Use it when you need the model to write or continue text. It is set up for gguf. The card lists the license as mit.
the-walking-fish/jijun-LM 之 GGUF 量化版本(v4d,2026-10-15)
Downloads · 30 days
708
60% of all-time downloads
All-time downloads
1.2K
Public
Repo size
1.9 GB
Likes
1
Public
Click a slice to open those files.
.gguf579 MB · 100%
From the Hugging Face model README
the-walking-fish/jijun-LM之 GGUF 量化版本(v4d,2026-10-15)
本 repo 目前為 v4d(2026-10-15),對應 the-walking-fish/jijun-LM 的 model-json/(同為 v4d)。
v4d 為 continue-SFT(v1 → v2 → v4b → v4c → v4d)最新一代,主打「使用者自訂分類名字面命中」修復:輸入句中出現的分類名若存在於 prompt 的 enum 內,即原樣抄寫(例:輸入 孝親費10000 即選 孝親費)。同時修復了 GGUF 分詞器前綴欄位缺失(見下方)。
| 檔案名稱 | 格式 (Type) | 體積 (Size) | 說明 |
|---|---|---|---|
bookkeeping_model_q4_0.gguf | Q4_0 | ~35.7 MB | 4-bit 量化:測試與網頁端側預設 |
bookkeeping_model_q5_0.gguf | Q5_0 | ~43.6 MB | 5-bit 量化 |
bookkeeping_model_q6_k.gguf | Q6_K | ~54.5 MB | 6-bit K-quant 量化 |
bookkeeping_model_q8_0.gguf | Q8_0 | ~67.2 MB | 8-bit 量化,接近 FP16 精度 |
bookkeeping_model_f16.gguf | FP16 | ~126.2 MB | Float16 原始精度 |
bookkeeping_model_f32.gguf | F32 | ~252.0 MB | Float32 全精度(驗證基準) |
本 repo 的 GGUF 一律寫入 tokenizer.ggml.pre = "olmo",與訓練端 HuggingFace tokenizer.json 的 ByteLevel 前分詞逐 token id 對齊(實測 80 筆多域 prompt 零差異)。
若使用不認得此欄位的自訂載入器,CJK 混排 prompt 會被 llama-bpe 字節前綴切分、偏離訓練分佈,推論品質會明顯劣化(自訂分類名場景全滅)。llama.cpp / wllama / llama-cpp-python 等正規實作會自動讀取並套用,無需處理。
輸入協定為裸 ChatML 拼接(無 HF chat template);輸出為 <|tool_call|>包裷的 JSON tool call。
llama-cpp-python (Python 測試)from llama_cpp import Llama
llm = Llama(
model_path="bookkeeping_model_q4_0.gguf",
n_ctx=2048,
n_threads=4,
verbose=False,
)
system_prompt = (
"今天是 2026-12-01(星期二)。你是一個記帳助理。你被賦予了以下 tools:\n"
'{"name": "add_record", "description": "新增一筆記帳記錄", "parameters": {"type": "object", "properties": {"amount": {"type": "number"}, "category": {"type": "string", "enum": ["飲食", "日常", "交通", "娛樂", "醫療", "教育", "還款", "薪水", "獎金", "零用錢", "兼職", "投資", "利息", "欠款回收", "其他"]}, "account": {"type": "string", "enum": ["現金", "信用卡", "悠遊卡", "一卡通", "街口支付", "LINE Pay", "Apple Pay", "Google Pay", "郵局帳戶", "銀行存款", "外幣帳戶", "加密貨幢", "悠遊付", "icash"]}, "description": {"type": "string"}, "type": {"type": "string", "enum": ["expense", "income"]}, "date": {"type": "string", "description": "ISO 8601 格式日期"}}, "required": ["amount", "category", "account", "type", "date"]}}'
)
prompt = (
"<|im_start|>system\n" + system_prompt + "<|im_end|>\n"
"<|im_start|>user\n早餐 60<|im_end|>\n"
"<|im_start|>assistant\n"
)
out = llm.create_completion(prompt=prompt, max_tokens=300, temperature=0.0,
stop=["<|tool_result|>", "<|im_end|>"])
print(out["choices"][0]["text"])
# 預期:<|tool_call|>{"name": "add_record", "args": {"amount": 60, "category": "飲食", "account": "現金", "description": "早餐", "type": "expense", "date": "2026-12-01"}}<|tool_result|>
@wllama/wllama (JavaScript / PWA)import { Wllama } from '@wllama/wllama';
const wllama = new Wllama({ 'llama.cpp': { suppressLog: true } });
await wllama.loadModelFromUrl(
'https://huggingface.co/the-walking-fish/jijun-LM-GGUF/resolve/main/bookkeeping_model_q4_0.gguf');
const response = await wllama.createCompletion(prompt, {
nPredict: 300,
sampling: { temp: 0 },
// stop 建議含標準 ChatML 與 tool 收尾標記
});
console.log(response);
評測集(831 筆分層語料,含日期擺點),持續改善與遺忘防線實測:
| 指標 | v4c | v4d |
|---|---|---|
| 格式通過率 | 100% | 100% |
| 金額 | 95.8% | 96.4% |
| 日期 | 77.1% | 80.6% |
| 收支型別 | 88.4% | 89.3% |
| 專案點名「孝親費10000」型場景(5 案,全 quant) | 0/5 | 5/5 |
MIT License。分詞器借用 jingyaogong/minimind-3 之 Llama 相容詞表並動態註冊記帳特殊 Token;模型權重為自訓。