Instructions to use lightonai/SASS-MiniCPM5-1B-Weighted-SFT-ColGREP with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use lightonai/SASS-MiniCPM5-1B-Weighted-SFT-ColGREP with Transformers:

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

pipe = pipeline("text-generation", model="lightonai/SASS-MiniCPM5-1B-Weighted-SFT-ColGREP")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("lightonai/SASS-MiniCPM5-1B-Weighted-SFT-ColGREP")
model = AutoModelForCausalLM.from_pretrained("lightonai/SASS-MiniCPM5-1B-Weighted-SFT-ColGREP", device_map="auto")
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=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use lightonai/SASS-MiniCPM5-1B-Weighted-SFT-ColGREP with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "lightonai/SASS-MiniCPM5-1B-Weighted-SFT-ColGREP"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "lightonai/SASS-MiniCPM5-1B-Weighted-SFT-ColGREP",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/lightonai/SASS-MiniCPM5-1B-Weighted-SFT-ColGREP

SGLang

How to use lightonai/SASS-MiniCPM5-1B-Weighted-SFT-ColGREP 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 "lightonai/SASS-MiniCPM5-1B-Weighted-SFT-ColGREP" \
    --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": "lightonai/SASS-MiniCPM5-1B-Weighted-SFT-ColGREP",
		"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 "lightonai/SASS-MiniCPM5-1B-Weighted-SFT-ColGREP" \
        --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": "lightonai/SASS-MiniCPM5-1B-Weighted-SFT-ColGREP",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use lightonai/SASS-MiniCPM5-1B-Weighted-SFT-ColGREP with Docker Model Runner:
```
docker model run hf.co/lightonai/SASS-MiniCPM5-1B-Weighted-SFT-ColGREP
```

SASS-MiniCPM5-1B-Weighted-SFT-ColGREP

Full model checkpoint for repository code localization using the colgrep harness.

Weighted supervised fine-tuning of openbmb/MiniCPM5-1B on 13,452 clean multilingual DeepSeek-V4-Flash-0731 trajectories, using configuration colgrep, split sft. Injected teacher finish reminders were removed from student training data.

Use the system prompt and tool schemas in student-protocol.json with the included tokenizer and chat template. The model requires an external harness to execute tool calls; repository snapshots and search indexes are not included. ColGREP models use semantic search; GREP models use lexical search through the terminal.

The original training configuration is included for provenance; local paths in that file must be adapted for another environment. provenance.json records checkpoint identity and SHA-256 hashes. Model weights are unchanged from the evaluated local checkpoint. Upstream base-model terms apply.

Consolidated data and reproducibility

Load lightonai/SASS-Data with configuration colgrep and split sft for weighted demonstrations or rl for RL tasks. The SFT tools column is JSON text; decode it before applying the chat template. Weights and tokenizer files are unchanged by the repository rename.

The packaged student protocol contains stricter search restrictions than Appendix A.6 of the paper. Preserve the supplied protocol for this artifact; reconciliation with the printed appendix remains outstanding.