Downloads · 30 days
18
75% of all-time downloads
SFT-4B-LiteOS
HaoranLiu/SFT-4B-LiteOS
SFT-4B-LiteOS is a image-text-to-text model from HaoranLiu. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for transformers.
Qwen/Qwen3-VL-4B-Instruct supervised-finetuned on Lite.OSWorld desktop computer-use trajectories, trained with cua-lite.
All-time downloads
24
Public
Parameters
4.4B
8.9 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors8.9 GB · 100%
At a glance
- Task
- Image-Text-to-Text
- Library
- transformers
- Model type
- qwen3_vl
- Access
- Public
- Created
- Aug 2, 2026
- Updated
- Aug 2, 2026
- SHA
- a0babc50
Try a prompt
Base models
Model card
From the Hugging Face model README
SFT-4B-LiteOS
Qwen/Qwen3-VL-4B-Instruct supervised-finetuned on Lite.OSWorld desktop
computer-use trajectories, trained with cua-lite.
The model takes a screenshot plus a task instruction and emits one Action: line
followed by <tool_call> blocks against a computer_use tool
(click / type / key / scroll / wait / terminate). It is trained with a 4-image
sliding history window at 1000x1000 screen resolution.
License is inherited from the base model; see Qwen/Qwen3-VL-4B-Instruct.
Training data
Exported from cua-lite/Lite.OSWorld
(desktop/use/train, the perturb + synth cohorts).
| Source trajectories | 2423 |
| Filter | not exclude_reason and episode_return > 0.5 |
| Kept | 1967 trajectories |
| Teacher | gpt-5.5 |
| Render config | scripts/configs/qwen3_vl/default/lite.osworld.yaml (4-image history, max_steps: 30) |
Training
Megatron backend via slime v0.3.0, 4x H100 80GB.
| Epochs | 3 (1475 rollouts, batch 4) |
| LR | 5e-6 cosine to 1e-6, warmup fraction 0.1 |
| Parallelism | TP=2, PP=1, CP=1, EP=1 -> DP=2 |
| Weight decay | 0.1 (Adam, betas 0.9/0.95) |
| Loss | sft_loss, per-token |
| Checkpoint | iter_1472 (final) |
Evaluation
lite.osworld eval split, greedy (temperature=0), max_steps: 30, concurrency 16.
The split holds 369 tasks; 37 carry an exclude_reason and are filtered out, leaving
332 evaluated tasks (all valid, group_size=1).
| Metric | Value |
|---|---|
Success rate (episode_return >= 1.0) | 104/332 = 31.3% |
| Mean episode return | 0.3231 |
Scores are near-binary: 224 zeros, 104 ones, and only 4 partial (0.536, 0.819, 0.903, 0.998).
| Domain | Success | Rate |
|---|---|---|
| thunderbird | 10/14 | 71.4% |
| vs_code | 11/18 | 61.1% |
| os | 10/19 | 52.6% |
| vlc | 7/15 | 46.7% |
| gimp | 7/16 | 43.8% |
| libreoffice_writer | 9/22 | 40.9% |
| chrome | 16/43 | 37.2% |
| libreoffice_impress | 16/47 | 34.0% |
| libreoffice_calc | 10/46 | 21.7% |
| multi_apps | 8/92 | 8.7% |
multi_apps is the largest bucket (28% of the eval set) and the weakest; excluding
it the remaining domains average 40.0%. No base-model baseline was run, so these
are absolute numbers rather than a measured delta.
Usage
from transformers import AutoProcessor, AutoModelForImageTextToText
model_id = "HaoranLiu/SFT-4B-LiteOS"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(model_id, device_map="auto")
Serving with SGLang:
python -m sglang.launch_server --model-path HaoranLiu/SFT-4B-LiteOS --port 30000
- Task
- Image-Text-to-Text
- Library
- transformers
- Type
- qwen3_vl
- Created
- Aug 2, 2026
- Updated
- Aug 2, 2026