Phi-4-mini FormosaNLU LoRA

這是 microsoft/Phi-4-mini-instruct 的 4-bit QLoRA adapter,用於正體中文口語 NLU: joint intent classification、slot filling,以及嚴格 JSON output。此 repository 不含 base model weights

它是 FormosaNLU-Synth 的第二個 student model family。相同 frozen paired contract 下, real + filtered synthetic 相較於 real only,在 Gemma 與 Phi 兩個 families 都得到正向、 且 hierarchical 95% CI 下界大於零的 intent accuracy 與 exact match 改善。

Artifact identity

欄位
Base model microsoft/Phi-4-mini-instruct
Frozen revision cfbefacb99257ffa30c83adab238a50856ac3083
Training arm real_syn_filtered
Seed 42
Adapter SHA-256 e9e4c77d79eb12da8cba64a7a484d260f753d000396752f51159e3c0f4e34376
Adapter type LoRA,r=16alpha=32、dropout 0.05
Target modules qkv_projo_projgate_up_projdown_proj
Steps 500
Public Dataset steven0226/formosa-nlu-synth-v1
Source kuotunyu/FormosaNLU-Synth
DOI 10.5281/zenodo.21879133 (v1.2.2)

這是一個可下載的 single adapter,不是六組 runs 的合併權重。release_manifest.json 記錄 adapter SHA-256、byte size、tensor count、base revision、training arm 與 source commit。上傳 bundle 僅允許 inference 所需檔案;optimizer state、checkpoints、 training_args.bin 與本機路徑不會發布。

Results

所有數字都來自未進入訓練的 frozen MASSIVE zh-TW Test(2,974 rows)。下表只描述本 repo 發布的 seed-42 adapter;跨 seeds 與跨 model-family 結論請以 source reports 為準。

Metric Real only Real + filtered synthetic Delta (pp)
Intent accuracy 72.49% 74.28% +1.78
Intent macro-F1 73.14% 74.02% +0.88
Slot micro-F1 59.55% 60.08% +0.53
Exact match 45.49% 46.70% +1.21
JSON-valid rate 96.67% 97.88% +1.21

三個 seeds(42/43/44)的 filtered-minus-real-only mean delta:intent accuracy +5.09 pp (hierarchical 95% CI [+1.83, +9.02]),exact match +4.71 pp[+1.36, +7.59])。完整 paired statistics 與 cross-family criterion:

Load the adapter

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

base_id = "microsoft/Phi-4-mini-instruct"
base_revision = "cfbefacb99257ffa30c83adab238a50856ac3083"
adapter_id = "steven0226/phi-4-mini-formosanlu-lora"

tokenizer = AutoTokenizer.from_pretrained(base_id, revision=base_revision)
model = AutoModelForCausalLM.from_pretrained(
    base_id,
    revision=base_revision,
    device_map="auto",
    torch_dtype=torch.bfloat16,
    quantization_config=BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_quant_type="nf4",
        bnb_4bit_use_double_quant=True,
        bnb_4bit_compute_dtype=torch.bfloat16,
    ),
)
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()

請沿用 source repository 的 formosanlu_nlu.v1 prompt 與 label catalog。這個 adapter 不是一般聊天 模型;若改變 prompt、label set、parser 或 decoding contract,以上指標不再可直接比較。

Limitations

  • 證據限於 MASSIVE zh-TW、一個 frozen split 與三個 training seeds。
  • 公開 artifact 是 seed 42;三-seed 結論來自獨立 runs,不代表任一單一 adapter 的保證。
  • Exact-match 對 JSON、intent label、slot type 與 span formatting 都很敏感。
  • Synthetic data 可能延續 teacher 或原始資料的偏誤;部署前仍需 domain-specific safety audit。
  • microsoft/Phi-4-mini-instruct 的 license、notices 與使用限制仍適用。

Citation

@software{kuotunyu_formosanlu_synth_2026,
  author  = {kuotunyu},
  title   = {FormosaNLU Synthetic Data Distillation for Traditional Chinese NLU},
  year    = {2026},
  version = {1.2.2},
  doi     = {10.5281/zenodo.21879133},
  url     = {https://doi.org/10.5281/zenodo.21879133}
}
Downloads last month
33
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for steven0226/phi-4-mini-formosanlu-lora

Adapter
(189)
this model

Dataset used to train steven0226/phi-4-mini-formosanlu-lora