---

language:

• zh

tags:

• chiboard
• pinyin
• ime
• sft
• dpo
• full-finetune
• gguf

base_model: johnbean393/chiboard-1-qwen3.5-0.8b-t1

base_model_relation: quantized

---

Chiboard 1 — Qwen3.5 0.8B

Chiboard 1 converts Chinese pinyin to Hanzi and supports revision of an existing

conversion.

Model lineage

johnbean393/chiboard-1-qwen3.5-0.8b-t1

was fine-tuned from Qwen/Qwen3.5-0.8B-Base with SFT and DPO using full

fine-tuning. This GGUF was quantized from that model.

Benchmarks

Final-checkpoint performance on a matched 5,000-row benchmark containing 2,500

plain conversion rows and 2,500 revision rows. Higher is better for exact match

(EM); lower is better for character error rate (CER).

| Checkpoint | Macro EM | Macro CER | Plain EM | Plain CER | Revision EM | Revision CER | Rewrite EM | Rewrite CER | Extension EM | Extension CER |

|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|

| Final | 63.68% | 0.0757 | 62.08% | 0.0981 | 65.28% | 0.0533 | 48.99% | 0.0711 | 85.50% | 0.0220 |

Download

Install the Hugging Face CLI,

then download the GGUF:

hf download johnbean393/chiboard-1-qwen3.5-0.8b-t1-GGUF \
  Chiboard-Qwen3.5-0.8B-T1-step36447-Q8_0.gguf \
  --local-dir .

Run with llama.cpp

Use llama-completion from a current

llama.cpp build. The prompt format is:

<|chiboard_context|>{committed_context}<|chiboard_pinyin|>{raw_pinyin}<|chiboard_display|>{display}<|chiboard_output|>

Fields may be empty. For example:

llama-completion \
  -m Chiboard-Qwen3.5-0.8B-T1-step36447-Q8_0.gguf \
  --no-conversation \
  --temp 0 \
  -n 128 \
  -p '<|chiboard_context|><|chiboard_pinyin|>wogeidashurourou<|chiboard_display|><|chiboard_output|>'