Iris1.5

IRIS1.5 is built on

  • ByteDance/Ouro-1.4B (language model)
  • NVlabs/RADIO radio_v2.5-h (vision model)

This repo does not include the full base models. It only includes the small trained parts from my run

  • projector
  • LoRA adapters

Those are stored as safetensors. Ouro and RADIO are downloaded automatically.

Contents

Bundle files

  • hf_bundle_step115000/projector.safetensors
  • hf_bundle_step115000/lora_edge/adapter_model.safetensors
  • hf_bundle_step115000/lora_gate/adapter_model.safetensors
  • hf_bundle_step115000/iris_bundle.json

Quick Start

1) Install dependencies

Install the requirements from this repository.

2) Run inference from this Hub repo

Example command from this codebase

python inference/infer_hf_bundle.py \
  --repo-id DavidSeyserHF/Iris1.5 \
  --image inference/complex.jpg \
  --prompt "What is in this image?"

Notes

  • First run downloads and caches base models.
  • If you want to force a specific bundle folder in the repo, use --bundle-subdir hf_bundle_step115000.

Training Setup (Short)

Base models

  • LLM is ByteDance/Ouro-1.4B
  • Vision model is NVlabs/RADIO radio_v2.5-h

Main config values from this checkpoint

  • image_size is 432
  • dtype is bf16
  • projector_hidden_mult is 1
  • projector_num_intermediate is 2
  • use_lora is true
  • lora_dual_edge_and_gate is true
  • lora_edge_layers is 4
  • lora_r is 16
  • lora_alpha is 32
  • lora_dropout is 0.05

Benchmarks

Benchmarking is currently work in progress.

Limitations and Biases

Known limitations

  • can be too verbose when short answers are needed
  • weak on strict answer formatting (yes/no/count/single word)

Safety and Responsible Use

Please do not use this model as the sole source for

  • medical, legal, financial, or safety-critical decisions
  • surveillance, profiling, or rights-impacting automated judgments

Human oversight and domain-specific validation are required for real-world deployments.

Reproducibility

Scripts used for this release

  • inference/export_hf_bundle.py
  • inference/infer_hf_bundle.py

License

This repository provides adapter/projector deltas only. Commercial use is restricted by the RADIO dependency license NVIDIA Source Code License-NC Link https://github.com/NVlabs/RADIO/blob/main/LICENSE Usage is also subject to the licenses and terms of

  • ByteDance/Ouro-1.4B
  • NVlabs/RADIO
  • any datasets used during training

Citation

If you use this release please cite this IRIS adapter release. Use the repository URL and commit hash for reproducibility.

@misc{seyser2026iris1-5,
  title = {Iris1.5},
  author = {David Seyser},
  year = {2026},
  howpublished = {Hugging Face model repository},
  url = {https://huggingface.co/DavidSeyserHF/Iris1.5},
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for DavidSeyserHF/Iris1.5

Adapter
(1)
this model

Datasets used to train DavidSeyserHF/Iris1.5

Space using DavidSeyserHF/Iris1.5 1

Collection including DavidSeyserHF/Iris1.5