TR-HASH-Vision-v8-2M-COCO-SFT

TR-HASH Vision v8 2M COCO SFT

This repository contains the best checkpoints from clean-image supervised refinement of TR-HASH Vision v8 2M COCO. The realized detector has 2.53M parameters and uses 640 px inputs.

Training was stopped during epoch 15/30 after the validation curve plateaued. The best O2M checkpoint was selected at epoch 12 by official COCO mAP50-95; the best NMS-free checkpoint was selected independently at epoch 13.

Evaluation

All values below come from the official COCO evaluator (faster backend) on COCO 2017 validation.

The root checkpoint is the best O2M + NMS model. The independently selected NMS-free checkpoint is under best_nms_free/. Exact evaluation payloads are in validation.json and validation_nms_free.json; metrics.jsonl contains the complete training trajectory.

Independent reproduction

The numbers above are flagged "self-reported" by the Hub's evaluation widget (standard for any model-index declaration not run through an official leaderboard integration) -- worth noting anyway that they have since been independently reproduced -- thanks to the community. Evaluation was rerun on this checkpoint and reported:

These match the values in the table above to within rounding.

Example detection

<p align="center"> <img alt="TR-HASH Vision v8 SFT detecting people, a tennis racket and a baseball glove during a baseball game" src="./assets/sft-example-baseball.png" width="500"> </p> <p align="center"><em>Qualitative O2M + NMS inference example at a 0.20 confidence threshold. This image is not part of the reported COCO evaluation.</em></p>

Comparison with the original YOLO26 reference

This keeps the same 32.2 AP YOLO26 reference used on the base checkpoint card, so the percentages remain directly comparable with the original result.

The SFT checkpoint improves the base checkpoint by +20.9% relative and reduces the remaining gap to the original YOLO26 reference from 48.5% to 37.7%. This is a progress comparison against the reference already used by the project, not an independently reproduced head-to-head benchmark; training recipes, initialization and compute budgets differ.

Inference

from PIL import Image
import torch

from complexity.generative.detection import (
    load_detector_from_hub,
    preprocess_detector_image,
    restore_detector_boxes,
)

device = "cuda" if torch.cuda.is_available() else "cpu"
model = load_detector_from_hub(
    "AETHORIA-AI/TR-HASH-Vision-v8-2M-COCO-SFT",
    device=device,
)
pixels, metadata = preprocess_detector_image(
    Image.open("image.jpg"), model.config.image_size
)

with torch.inference_mode():
    prediction = model.predict(pixels[None].to(device))[0]
prediction["boxes"] = restore_detector_boxes(
    prediction["boxes"].cpu(), metadata
)

Class IDs are listed in class_names.json. Boxes returned by predict are normalized xyxy coordinates until restore_detector_boxes maps them to source pixels.

Training

• Dataset: COCO 2017 train; evaluation on COCO 2017 validation
• Initialization: detector transfer from AETHORIA-AI/TR-HASH-Vision-v8-2M-COCO
• Purpose: full-parameter supervised vision refinement
• Optimizer: MuSGD
• Input: clean-image/light augmentation, multi-resolution 512-640 px
• EMA: 0.9999
• Checkpoint selection: official COCO mAP50-95, independently per inference branch
• Framework: Complexity Framework

Limitations

This is a research checkpoint under CC BY-NC 4.0. Validate accuracy, calibration, latency, and failure modes on the target domain before deployment.