Instructions to use SC2meisah/BioCompute-ARB-Qwen3-8B-GRPO with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use SC2meisah/BioCompute-ARB-Qwen3-8B-GRPO with Transformers:

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

pipe = pipeline("text-generation", model="SC2meisah/BioCompute-ARB-Qwen3-8B-GRPO")
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("SC2meisah/BioCompute-ARB-Qwen3-8B-GRPO")
model = AutoModelForCausalLM.from_pretrained("SC2meisah/BioCompute-ARB-Qwen3-8B-GRPO", 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 SC2meisah/BioCompute-ARB-Qwen3-8B-GRPO with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "SC2meisah/BioCompute-ARB-Qwen3-8B-GRPO"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "SC2meisah/BioCompute-ARB-Qwen3-8B-GRPO",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/SC2meisah/BioCompute-ARB-Qwen3-8B-GRPO

SGLang

How to use SC2meisah/BioCompute-ARB-Qwen3-8B-GRPO 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 "SC2meisah/BioCompute-ARB-Qwen3-8B-GRPO" \
    --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": "SC2meisah/BioCompute-ARB-Qwen3-8B-GRPO",
		"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 "SC2meisah/BioCompute-ARB-Qwen3-8B-GRPO" \
        --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": "SC2meisah/BioCompute-ARB-Qwen3-8B-GRPO",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use SC2meisah/BioCompute-ARB-Qwen3-8B-GRPO with Docker Model Runner:
```
docker model run hf.co/SC2meisah/BioCompute-ARB-Qwen3-8B-GRPO
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

BioCompute-ARB-Qwen3-8B-GRPO

Qwen3-8B trained to act as a tool-using research agent on BioCompute-ARB, a small Auto Research Benchmark for scientific computing in computational biology (pairwise alignment with affine gaps, stiff reaction-network simulation, SEIR change-point inference). Each task is graded by an independent program on hidden test cases.

Code, benchmark tasks, training scripts and evaluation logs: https://github.com/sc2musa/BioCompute-ARB-Qwen3-8B-GRPO

Training

Rejection-sampling fine-tuning (RFT). Qwen3-8B generated agent trajectories on freshly generated training instances (new seeds; the benchmark's hidden test cases were never used). Trajectories were sampled with method guidance, scored by the task grader, and the kept ones were trained on with the goal-only instruction (guidance removed). Full-parameter, 3587 turn samples from the reaction-network task, 2 epochs, lr 1e-5.
Multi-turn agentic GRPO with ROLL, starting from the RFT model: 4 attempts per training instance, turn-level samples sharing the episode return (score + 0.5 if solved), group-normalised advantages, no KL, lr 1e-6, 32 trajectories per step, 60 steps on 256 alignment and reaction-network training instances. This is the final checkpoint.

No teacher model or human-written solutions were used.

Results (held-out test set, 20 trials per task)

Configuration F: goal-only instruction, non-thinking mode, repeat guard, public self-check script.

Model	Reaction network: solved / mean score	Alignment: mean score	SEIR: mean score
Qwen3-8B	0/20 / 0.656	0.154	0.031
+ RFT	11/20 / 0.550	0.047	0.000
+ RFT + GRPO (this model)	19/20 / 0.950	0.068	0.000

Usage

Use non-thinking mode (enable_thinking=False) with tool calling, e.g. vLLM:

vllm serve SC2meisah/BioCompute-ARB-Qwen3-8B-GRPO --enable-auto-tool-choice \
  --tool-call-parser hermes --reasoning-parser qwen3 --max-model-len 40960

and the agent harness of the benchmark repository (harness/runner.py --agent llm --guidance goal --repeat-guard --public-check), sampling with temperature 0.7, top_p 0.8.

Limitations

The improvement is concentrated on the reaction-network task, where the base model already succeeded occasionally. No model solves the alignment or SEIR tasks; the alignment score is below the base model.
With thinking mode enabled the model often writes the program inside its reasoning and submits without writing the file (10/20 on the reaction-network task instead of 19/20).
Evaluated with 20 trials per task on one benchmark; test instances come from the same generators as the training instances (different seeds), so out-of-distribution generalisation is untested.

Downloads last month: 31

Safetensors

Model size

8B params

Tensor type

BF16

Model tree for SC2meisah/BioCompute-ARB-Qwen3-8B-GRPO

Base model

Qwen/Qwen3-8B-Base

Finetuned

Qwen/Qwen3-8B

Finetuned

(2238)

this model

Quantizations

1 model