Instructions to use ynanxiu/olmo3-190M-zh-full with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ynanxiu/olmo3-190M-zh-full with Transformers:

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

pipe = pipeline("text-generation", model="ynanxiu/olmo3-190M-zh-full")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("ynanxiu/olmo3-190M-zh-full")
model = AutoModelForCausalLM.from_pretrained("ynanxiu/olmo3-190M-zh-full")
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 ynanxiu/olmo3-190M-zh-full with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ynanxiu/olmo3-190M-zh-full"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ynanxiu/olmo3-190M-zh-full",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ynanxiu/olmo3-190M-zh-full

SGLang

How to use ynanxiu/olmo3-190M-zh-full 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 "ynanxiu/olmo3-190M-zh-full" \
    --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": "ynanxiu/olmo3-190M-zh-full",
		"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 "ynanxiu/olmo3-190M-zh-full" \
        --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": "ynanxiu/olmo3-190M-zh-full",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use ynanxiu/olmo3-190M-zh-full with Docker Model Runner:
```
docker model run hf.co/ynanxiu/olmo3-190M-zh-full
```

OLMo3-190M-zh-full

从零开始、在 RTX 4060 8GB 上预训练的 190M 中文语言模型。

模型概述

这是一个基于 OLMo3 架构的 190M 参数中文预训练模型，完全从随机初始化（from scratch）开始训练，未依赖任何已有预训练权重。

🏗️ 架构: OLMo3（Sliding Window Attention + Full Attention 混合）
📏 参数: 187M 总参数（113M 非嵌入参数）
🌐 语言: 中文
📚 训练数据: cmz1024/llm101-olmo3-zh-demo-data（~93 万条中文序列）
🖥️ 硬件: NVIDIA RTX 4060 8GB（本地训练）
⏱️ 训练耗时: ~47 小时

模型架构

参数	值
hidden_size	768
num_layers	12
num_heads	12
intermediate_size	3072
vocab_size	48,000
max_position_embeddings	2048
注意力机制	Sliding Window（每 4 层中 3 层）+ Full Attention（1 层）

训练细节

配置	值
优化器	AdamW (β1=0.9, β2=0.95)
学习率	5e-4（cosine decay）
Warmup	2%（~320 steps）
序列长度	2048
Batch Size	1 × 128 grad_accum（等效 128）
精度	bf16
总步数	7,675
Eval Loss	3.624

Loss 曲线

Step   500:  eval_loss=5.307
Step  1000:  eval_loss=4.534
Step  1500:  eval_loss=4.243
Step  2000:  eval_loss=4.087
Step  2500:  eval_loss=3.986
Step  3000:  eval_loss=3.906
Step  3500:  eval_loss=3.841
Step  4000:  eval_loss=3.790
Step  4500:  eval_loss=3.747
Step  5000:  eval_loss=3.709
Step  5500:  eval_loss=3.678
Step  6000:  eval_loss=3.654
Step  6500:  eval_loss=3.637
Step  7000:  eval_loss=3.627
Step  7500:  eval_loss=3.624

初始 loss 10.92 → 最终 3.624，**下降 66.8%**。

使用方法

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "ynanxiu/olmo3-190M-zh-full",
    torch_dtype=torch.bfloat16,
    attn_implementation="sdpa"
).to("cuda")

tokenizer = AutoTokenizer.from_pretrained("ynanxiu/olmo3-190M-zh-full")

prompt = "从前有座山，山里有座庙，"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

outputs = model.generate(
    **inputs,
    max_new_tokens=100,
    do_sample=True,
    temperature=0.8,
    top_p=0.9,
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

生成示例

Prompt: 从前有座山，山里有座庙，

生成: 庙里有一座庙宇，庙里有一棵大树。有一天早晨起来，发现一棵树上有个苹果树...两个大猴子在树下各捉一个，猴子在树上捉一个...

Prompt: 人工智能是

生成: 人工智能是未来的趋势，它给人类带来了新的机会。在人工智能领域，人工智能的潜力是非常大的...

限制与用途

✅ 适用于：中文文本生成、续写、作为下游任务的基座模型
⚠️ 注意：这是基础预训练模型，未经过 SFT/RLHF，输出为续写风格而非对话风格
⚠️ 模型可能产生重复或不一致的内容（190M 参数规模限制）
🔧 建议：可在此基础上进行持续预训练或 SFT 以获得更好效果

环境足迹

硬件：NVIDIA RTX 4060 8GB（消费级 GPU）
训练功耗：~115W × 47h ≈ 5.4 kWh
训练方式：本地训练，无云计算碳排放

Model tree for ynanxiu/olmo3-190M-zh-full

Adapters

1 model

Finetunes