Instructions to use OraRL/Video-ORA-9B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use OraRL/Video-ORA-9B with Transformers:

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

pipe = pipeline("image-text-to-text", model="OraRL/Video-ORA-9B")
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("OraRL/Video-ORA-9B")
model = AutoModelForMultimodalLM.from_pretrained("OraRL/Video-ORA-9B", 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 OraRL/Video-ORA-9B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "OraRL/Video-ORA-9B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "OraRL/Video-ORA-9B",
		"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/OraRL/Video-ORA-9B

SGLang

How to use OraRL/Video-ORA-9B 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 "OraRL/Video-ORA-9B" \
    --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": "OraRL/Video-ORA-9B",
		"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 "OraRL/Video-ORA-9B" \
        --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": "OraRL/Video-ORA-9B",
		"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 OraRL/Video-ORA-9B with Docker Model Runner:
```
docker model run hf.co/OraRL/Video-ORA-9B
```

Video-ORA-9B / code /eval /README.md

lyhisme

Add OraRL training and evaluation code

53c10a4 verified about 17 hours ago

preview code

Raw

History Blame Contribute Delete

3.31 kB

	# Evaluation runtime

	`eval/task/` holds the evaluators that produce every number in the paper. They
	ship with this repository, so a single clone can reproduce the full suite.
	`orarl-eval` is the supported entry point: it resolves the canonical dataset
	layout into per-task environment variables and then runs `eval/task/eval.sh`.

	```bash
	orarl-eval \
	--model /path/to/exported-model \
	--tasks paper \
	--dataset OraRL/OraRL-Data \
	--summary ./evaluation.json \
	--run
	```

	Without `--run` the command prints the resolved plan and exits, which is the
	fastest way to confirm data and model paths before allocating GPUs.
	`eval/task/eval.sh` can also be called directly when you want to bypass the
	canonical dataset resolution and set the per-task variables yourself.

	## Task map

	\| Family \| Benchmarks \| Entry point \|
	\| --- \| --- \| --- \|
	\| Video QA \| VideoMME, VideoMME-v2, MVBench, MMVU, Video-Holmes, LongVideoBench, MLVU \| `eval_vllm.py` \|
	\| Spatial intelligence \| VSI-Bench, MindCube, MMSI-Bench, ReVSI \| `eval_vllm.py`, `mmsi/eval_mmsi_transformers.py`, `revsi/eval_revsi_vllm.py` \|
	\| Temporal grounding \| TimeLens (Charades, ActivityNet, QVHighlights) \| `temporal_grounding/eval_timelens_hf.py` \|
	\| Spatial grounding \| RefCOCO, RefCOCO+, RefCOCOg \| `spatial_grounding/eval_refcoco_vllm.py` \|
	\| Tracking \| GOT-10k \| `tracking/eval_tracking_vllm.py` \|
	\| Spatio-temporal grounding \| STVG \| `spatial_temporal_grounding/eval_stvg_vllm.py` \|
	\| Segmentation \| RefCOCO series, MeViS, ReasonVOS \| `segmentation/eval_seg_vllm.py` plus `segmentation/post_sam2.py` \|

	`eval_prompt.py` is the single source of truth for prompts, and
	`canonical_data.py` adapts the canonical `ORARL_EVAL_*` layout for every
	evaluator. MMSI-Bench runs through Transformers rather than vLLM; the other
	families run through vLLM.

	## Assets you must obtain separately

	Annotations and media come from the `OraRL/OraRL-Data` dataset repository, and
	model weights come from the released checkpoints. Neither is vendored here.

	Segmentation additionally needs three inputs, all supplied by you under their
	upstream licenses:

	- SAM2 weights and the matching Hydra config, passed as
	`SEGMENTATION_SAM2_CKPT` and `SEGMENTATION_SAM2_CFG`.
	- The official OneThinker `seg_post_sam2.py`, passed as
	`SEGMENTATION_POSTPROCESSOR_PATH`. `segmentation/post_sam2.py` is a thin
	wrapper that injects paths into it so mask metrics stay identical to the
	upstream implementation.
	- The `sam2` Python package.

	`orarl-eval` refuses to start segmentation with `--segmentation-run-sam2`
	unless all three paths exist.

	## Video decoding

	Video evaluators pin the decord backend through `FORCE_QWENVL_VIDEO_READER`.
	`eval/task/qwenvl_decord_patch.py` replaces `qwen_vl_utils.fetch_video` so the
	backend is honoured across `qwen_vl_utils` releases that otherwise hard-code
	torchvision. The backend is not cosmetic: switching MeViS and ReasonVOS from
	torchcodec to decord moved MeViS J&F from 56.7 to 60.6. Set
	`SEGMENTATION_VIDEO_READER` to `torchcodec` or `torchvision` only when
	deliberately measuring that difference.

	## Checkpoint format

	Exported Hugging Face checkpoints run as-is. Pass `--force-merge` when the model
	path is a sharded FSDP actor directory; `eval/task/eval.sh` then merges it with
	`scripts/model_merger.py` before inference.