Instructions to use HyperSol/ResearchPhysics-9B-SFT with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use HyperSol/ResearchPhysics-9B-SFT with Transformers:

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

pipe = pipeline("image-text-to-text", model="HyperSol/ResearchPhysics-9B-SFT")
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)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("HyperSol/ResearchPhysics-9B-SFT")
model = AutoModelForMultimodalLM.from_pretrained("HyperSol/ResearchPhysics-9B-SFT", 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=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use HyperSol/ResearchPhysics-9B-SFT with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "HyperSol/ResearchPhysics-9B-SFT"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "HyperSol/ResearchPhysics-9B-SFT",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/HyperSol/ResearchPhysics-9B-SFT

SGLang

How to use HyperSol/ResearchPhysics-9B-SFT 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 "HyperSol/ResearchPhysics-9B-SFT" \
    --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": "HyperSol/ResearchPhysics-9B-SFT",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

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 "HyperSol/ResearchPhysics-9B-SFT" \
        --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": "HyperSol/ResearchPhysics-9B-SFT",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use HyperSol/ResearchPhysics-9B-SFT with Docker Model Runner:
```
docker model run hf.co/HyperSol/ResearchPhysics-9B-SFT
```

ResearchPhysics-9B-SFT — E12 / step 1572

实验中 / Experimental — Training in progress as a project stage; no validated model release. This specific SFT run completed 16 epochs. The uploaded files are its epoch-12 checkpoint, selected using the development FSR evaluation. This project-stage label does not claim that a GPU training job is currently running.

Checkpoint and training

Item	Value
Base	Qwen/Qwen3.5-9B
Immutable base revision	`c202236235762e1c871ad0ccb60c8ee5ba337b9a`
Run	`20261104_sft9b_selected_plus_436_610_e16_v1`
Uploaded checkpoint	`checkpoint-1572`, epoch 12 of a planned 16
Data	4,186 private K3 original/registered repaired trajectories
Method	Full SFT; frozen vision tower and aligner; thinking enabled
Hardware / batch	4×H200 / global batch 32
Learning rate	1e-5, cosine decay over 16 epochs, 5% warmup
Sequence limit / seeds	131,072 tokens / seed 42, data seed 42
Supervision	Assistant Thinking + Final + EOS; user prompt masked; no rubric input
Validation	No held-out validation split

Exact data/config hashes and every uploaded weight/loading-file hash are in RELEASE_MANIFEST.json. Training datasets remain private and are not included. The export Git commit records preparation state, not a verified original training-code commit.

Development FSR evaluation

20 input questions, two generations per input, 40/40 scored; average trials within each input and then weight all 20 inputs equally. The score mean is 2.9228125/10; the fraction of responses scoring ≥7 is 10% (4/40). Generation used thinking with a 131,072-token output budget. Scoring used the existing FSR scorer with gpt-5.6-sol and the frozen rubric/prompt.

This benchmark informed dataset development and checkpoint selection. These are development measurements, not independent held-out certification, pass@2, or proof of a causal training benefit. Outputs and Judge grades can be incorrect. Historical 9B results are separate from the matched 27B main experiment.

Loading and limitations

The repository includes the unchanged checkpoint safetensors, model config, tokenizer, chat template and processor files. It uses the Qwen3.5 conditional- generation architecture; the training environment recorded Transformers 5.16.1. Use a runtime supporting that architecture and the supplied chat template with thinking enabled. The exact source checkpoint passed native saving and FSR inference; an additional GPU smoke test of this public export has not been run. The 128k training limit does not certify long-context or multimodal capability.

Physics outputs require independent checks. Training-set token accuracy, a valid file export, and development benchmark scores do not establish scientific correctness or generalization to unseen papers and tasks.

Licensing and provenance

Project-owned SFT modifications and new repository documentation are offered under CC BY-NC-ND 4.0, only to the extent of rights held by the project. The base-model components retain Apache-2.0; see UPSTREAM_LICENSE_APACHE_2.0.txt. The project license does not replace upstream grants or license private data. NOTICE.md identifies the checkpoint, modified files and scope. Earlier repository revisions and their valid permissions remain in history.

Downloads last month: 1

Safetensors

Model size

1.47M params

Tensor type

BF16

Model tree for HyperSol/ResearchPhysics-9B-SFT

Base model

Qwen/Qwen3.5-9B-Base

Finetuned

Qwen/Qwen3.5-9B

Finetuned

(1,299)

this model