Instructions to use johnbean393/chiboard-1-m2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use johnbean393/chiboard-1-m2 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="johnbean393/chiboard-1-m2")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("johnbean393/chiboard-1-m2")
model = AutoModelForCausalLM.from_pretrained("johnbean393/chiboard-1-m2")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use johnbean393/chiboard-1-m2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "johnbean393/chiboard-1-m2"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "johnbean393/chiboard-1-m2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/johnbean393/chiboard-1-m2

SGLang

How to use johnbean393/chiboard-1-m2 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "johnbean393/chiboard-1-m2" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "johnbean393/chiboard-1-m2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "johnbean393/chiboard-1-m2" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "johnbean393/chiboard-1-m2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use johnbean393/chiboard-1-m2 with Docker Model Runner:
```
docker model run hf.co/johnbean393/chiboard-1-m2
```

Chiboard-1 M2

Final 350M Chiboard M2 model for Chinese pinyin-to-Hanzi conversion and revision. It was initialized from johnbean393/chiboard-1-s1 and trained with correctness-gated, off-policy distillation from johnbean393/chiboard-1-t1.

This is the final M2 artifact and has completed the full validation protocol: one million training examples, integrity checks, and four frozen 140K SGLang evaluations. Validation-complete does not mean statistically proven superior to S1: its final point estimate is directionally better, but the predeclared 95% confidence interval includes zero and the superiority gate did not pass.

Prompt format

<|startoftext|>{committed_context}<|reserved_6|>{raw_pinyin}<|reserved_7|>{display}<|reserved_8|>{target}<|im_end|>

The tokenizer adds exactly one BOS. At inference time, stop the prompt after <|reserved_8|> and greedily generate through <|im_end|>.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "johnbean393/chiboard-1-m2"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

committed_context = ""
raw_pinyin = "nihao"
display = ""
prompt = (
    f"{committed_context}<|reserved_6|>{raw_pinyin}"
    f"<|reserved_7|>{display}<|reserved_8|>"
)
inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=True).to(model.device)
outputs = model.generate(
    **inputs,
    do_sample=False,
    max_new_tokens=96,
    eos_token_id=tokenizer.convert_tokens_to_ids("<|im_end|>"),
    pad_token_id=tokenizer.pad_token_id,
)
completion = outputs[0, inputs.input_ids.shape[1]:]
print(tokenizer.decode(completion, skip_special_tokens=True))

For a pure-pinyin active buffer, display is empty. For revision use, display contains the currently displayed Hanzi text to be corrected.

Training

Student initialization and frozen retention reference: johnbean393/chiboard-1-s1 at ac4e8c3885ef1a0ad5b3f90a1e4c5639ae669edd.
Teacher: johnbean393/chiboard-1-t1 at 4086865d8813a01909a420579ee7b15821bf80b0.
Exposure: 500,000 plain + 500,000 revision examples, balanced and document-disjoint from the training shadow evaluation.
Gold NLL on every completion token, including EOS.
Teacher candidate-set forward KL only where T1's top token is gold and frozen S1's top token is not gold.
Frozen-S1 retention KL on S1-correct tokens.
Candidate/retention top-k 8; teacher temperature 2.0; teacher weight 1.0; retention temperature 1.0; retention weight 0.1.
Peak LR 3e-7, 2% warmup, cosine decay to 3e-8 over 30,000 examples, then held at 3e-8 through 1,000,000 examples.
7,810 optimizer updates; three isolated clipped updates; no consecutive clipping or numerical instability.

The complete sanitized configuration and lineage are in training_summary.json.

Chiboard evaluation

The model was evaluated using the frozen johnbean393/chiboard-1-eval campaign sample: 100,000 plain and 40,000 revision rows, seed 20260714. Decoding used SGLang 0.5.9, BF16, greedy generation, the native model implementation, default FA3 attention, and shuffled request order.

Cohort	Exact match	CER	Delta EM vs S1	Delta CER vs S1
Plain	54.9171%	0.133746	+0.0757 pp	-0.0350 pp
Revision	54.6658%	0.076984	-0.0134 pp	-0.0432 pp
Hard ambiguity	52.0561%	0.139144	+0.0839 pp	-0.0337 pp

Combined utility versus S1 was +0.258 pp, with a 95% paired document-bootstrap CI of [-0.019, +0.546] pp. The gate verdict was FAIL because its lower bound did not exceed zero. There were no non-inferiority guard violations.

The exact SGLang manifest, report, and gate verdict are included under evaluation/.

Weights

Format: safetensors
Tensor file size: 708,984,464 bytes
SHA-256: 427331f979b1909bf400e94e43828bca397e97d03a0b68e103d3b2dd07ca85b8

Training telemetry: W&B run.

Downloads last month: 108

Safetensors

Model size

0.4B params

Tensor type

BF16

Model tree for johnbean393/chiboard-1-m2

Base model

LiquidAI/LFM2.5-350M-Base

Finetuned

johnbean393/chiboard-1-s1