How to save model predictions to JSON or CSV
Save the detections from RF-DETR, YOLO, Transformers, Detectron2 or any other model to a CSV or JSON file with supervision.
Overview
Saving a model's predictions to a file lets you analyze them in a spreadsheet or pandas, compare two runs, or hand them to another system. supervision (opens in a new tab), Roboflow's open source computer vision library, has two sinks that write detections to a file: sv.CSVSink and sv.JSONSink. Both take an sv.Detections object, so they work with any model supervision can read.
This guide covers:
- Saving RF-DETR predictions to CSV and JSON in Python.
- Saving predictions from other models: Ultralytics YOLO, YOLOv5, Transformers, Detectron2, MMDetection, YOLO-NAS, PaddleDetection, TensorFlow and DeepSparse.
- Saving predictions for every frame of a video.
- Asking a coding agent to do it through the Roboflow MCP server.
Step 1: Install supervision and RF-DETR
RF-DETR (opens in a new tab) is Roboflow's open source real-time model for object detection and segmentation. Its Nano, Small, Medium and Large checkpoints are licensed under Apache 2.0, so you can use them in commercial products. Install both packages:
pip install rfdetr supervisionStep 2: Save predictions to CSV
This script runs RF-DETR on every image in a folder and writes one row per detected object to predictions.csv. custom_data adds a column with the file name, so you can tell which image each row came from:
from pathlib import Path
import supervision as sv
from rfdetr import RFDETRSmall
model = RFDETRSmall()
with sv.CSVSink("predictions.csv") as sink:
for path in sorted(Path("images").glob("*.jpg")):
detections = model.predict(str(path), threshold=0.5)
sink.append(detections, custom_data={"image": path.name})Every row starts with the box corners in pixels (x_min, y_min, x_max, y_max), then class_id, confidence and tracker_id, which stays empty unless you track objects. The extra columns follow in alphabetical order: here class_name, image and source_shape. RF-DETR adds class_name and source_shape (the image height and width), and custom_data adds image. An image with no detections writes no rows.
RFDETRSmall() is pre-trained on the 80 COCO classes. To save predictions from your own model, load the checkpoint you fine-tuned with the same size class, for example RFDETRSmall(pretrain_weights="checkpoint_best_total.pth").
Step 3: Save predictions to JSON
Swap the sink for sv.JSONSink to write the same fields as JSON:
from pathlib import Path
import supervision as sv
from rfdetr import RFDETRSmall
model = RFDETRSmall()
with sv.JSONSink("predictions.json") as sink:
for path in sorted(Path("images").glob("*.jpg")):
detections = model.predict(str(path), threshold=0.5)
sink.append(detections, custom_data={"image": path.name})The file is one list with an object per detection, keyed like the CSV columns. Coordinates and confidence are numbers, and class_id is an integer. JSONSink keeps the rows in memory and writes the file when the with block ends.
Save predictions from other models
The sinks only need an sv.Detections object. For any other model, convert its output with the matching loader, then call sink.append() the same way:
- Ultralytics YOLO (YOLOv8, YOLO11, YOLO26):
sv.Detections.from_ultralytics() - YOLOv5:
sv.Detections.from_yolov5() - Hugging Face Transformers:
sv.Detections.from_transformers() - Detectron2:
sv.Detections.from_detectron2() - MMDetection:
sv.Detections.from_mmdetection() - YOLO-NAS:
sv.Detections.from_yolo_nas() - PaddleDetection:
sv.Detections.from_paddledet() - TensorFlow Hub:
sv.Detections.from_tensorflow() - DeepSparse:
sv.Detections.from_deepsparse() - Roboflow Inference and the hosted API:
sv.Detections.from_inference()
For example, with an Ultralytics model:
import supervision as sv
from ultralytics import YOLO
model = YOLO("yolo11n.pt")
with sv.CSVSink("predictions.csv") as sink:
result = model("image.jpg")[0]
detections = sv.Detections.from_ultralytics(result)
sink.append(detections, custom_data={"image": "image.jpg"})Save predictions from a video
For a video, read it frame by frame and record the frame number with each row:
import supervision as sv
from rfdetr import RFDETRSmall
model = RFDETRSmall()
frames = sv.get_video_frames_generator("video.mp4")
with sv.CSVSink("video-predictions.csv") as sink:
for index, frame in enumerate(frames):
detections = model.predict(frame[:, :, ::-1], threshold=0.5)
sink.append(detections, custom_data={"frame": index})Frames come out of OpenCV in BGR order and RF-DETR expects RGB, so frame[:, :, ::-1] flips the channels. How to run a computer vision model on video covers drawing the predictions on the video as well.
Save predictions with a coding agent
The Roboflow MCP server connects a coding agent, such as Claude Code, Cursor or Codex, to your Roboflow workspace. Describe the file you want and the agent writes and runs the script above with your model. For example:
Run RF-DETR Small on every image in ./images and save the detections to predictions.csv, one row per box with the image file name.