Instructions to use Natarizki/CMLM-0.8B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Natarizki/CMLM-0.8B with Transformers:

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

pipe = pipeline("text-generation", model="Natarizki/CMLM-0.8B")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("Natarizki/CMLM-0.8B")
model = AutoModelForMultimodalLM.from_pretrained("Natarizki/CMLM-0.8B", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.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(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Natarizki/CMLM-0.8B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Natarizki/CMLM-0.8B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Natarizki/CMLM-0.8B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Natarizki/CMLM-0.8B

SGLang

How to use Natarizki/CMLM-0.8B 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 "Natarizki/CMLM-0.8B" \
    --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": "Natarizki/CMLM-0.8B",
		"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 "Natarizki/CMLM-0.8B" \
        --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": "Natarizki/CMLM-0.8B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Natarizki/CMLM-0.8B with Docker Model Runner:
```
docker model run hf.co/Natarizki/CMLM-0.8B
```

CMLM-0.8B

Coding + Math Language Model — a Qwen3.5-0.8B fine-tuned for code generation and mathematical reasoning.

Model Details

Property	Value
Base Model	unsloth/Qwen3.5-0.8B
Architecture	Qwen3.5 (Gated DeltaNet + Full Attention hybrid)
Parameters	0.8B
Training Method	LoRA (r=16, α=32)
Precision	float32 (no quantization)
Max Context	2048 tokens
Framework	Unsloth + TRL SFTTrainer
Hardware	NVIDIA Tesla T4 (16 GB VRAM)
Training Steps	4,000

Training Data

Dataset	Samples	Domain
Magicoder-Evol-Instruct-110K	110,000	Code instruction following
MetaMathQA	395,000	Mathematical reasoning
NuminaMath-CoT	860,000	Math chain-of-thought
riddles_v1	469	Riddles & logical reasoning
Total	~1,366,000

⚠️ Note: Riddle dataset represents only 0.03% of total training data. Riddle/logic puzzle performance remains limited due to insufficient exposure. See Limitations below.

Training Hyperparameters

learning_rate: 2e-4
max_steps: 4000
per_device_train_batch_size: 4
gradient_accumulation_steps: 2
effective_batch_size: 16
warmup_steps: 100
optimizer: adamw_torch_fused
gradient_checkpointing: unsloth
lora_r: 16
lora_alpha: 32
lora_dropout: 0
target_modules: [q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj]
packing: true (ignored for processor-based model)
max_seq_length: 2048
precision: float32

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Natarizki/CMLM-0.8B", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("Natarizki/CMLM-0.8B")

messages = [{"role": "user", "content": "Write a Python function to check if a number is prime."}]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(model.device)

outputs = model.generate(inputs, max_new_tokens=256, temperature=0.7, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Benchmarks

Benchmarked on NVIDIA T4 with float32 inference via Unsloth. Results are non-deterministic (temperature=0.7).

Throughput

Domain	CMLM-0.8B (tok/s)	Base Qwen3.5-0.8B (tok/s)	Avg Latency (CMLM)
Coding	12.1	15.2	39.0s
Math	14.8	15.1	15.9s
General	15.0	15.1	17.0s

Lower coding tok/s reflects longer, more detailed code responses — not degraded quality.

Qualitative Assessment

Domain	Status	Notes
Code Generation	✅ Strong	Correct Python with docstrings, clean style
Math Reasoning	✅ Good	Step-by-step CoT, correct answers on standard problems
Definitions/Explanations	✅ Accurate	Concise, technically correct
Arithmetic	⚠️ Variable	Correct ~50% at temp=0.7; verify critical outputs
Riddles/Logic Puzzles	❌ Weak	Insufficient training signal (0.03% of data)

Limitations

Riddles and linguistic ambiguity: Model fails on trick questions (e.g., "all but 3 die") due to only 469 riddle samples in 1.36M total training data. A dedicated second-stage fine-tune on riddle/logic datasets is recommended for improvement.
Arithmetic consistency: Multi-step arithmetic accuracy varies across generations at temperature > 0. Use temperature=0.0 for deterministic outputs when correctness matters.
Training scope: 4,000 steps covers ~4.7% of the full dataset. Additional training would improve coverage and consistency.
No vision capabilities: Despite Qwen3.5's native multimodal architecture, this model was trained text-only.
Context limit: 2048 tokens maximum; not suitable for long-document tasks.

License

Apache 2.0 (inherits from Qwen3.5)

Acknowledgments

Qwen Team for the base model
Unsloth for efficient T4-compatible training
Dataset authors: Magicoder, MetaMath, NuminaMath, and Hypersniper teams

Downloads last month: 397

Safetensors

Model size

0.9B params

Tensor type

F32

BF16

Model tree for Natarizki/CMLM-0.8B

Base model

Qwen/Qwen3.5-0.8B-Base

Finetuned

Qwen/Qwen3.5-0.8B

Finetuned

unsloth/Qwen3.5-0.8B

Adapter

(39)

this model

Adapters

1 model

Natarizki
/

CMLM-0.8B