Downloads · 30 days
3
9% of all-time downloads
ortho-vision-qlora-final-v4
hiyasvyas/ortho-vision-qlora-final-v4
ortho-vision-qlora-final-v4 is a machine learning model from hiyasvyas. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for peft. The card lists the license as mit.
Final v4 LoRA adapter for orthographic projection behavior.
All-time downloads
35
Public
Repo size
340 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors329 MB · 97%
At a glance
- Library
- peft
- License
- mit
- Access
- Public
- Created
- Jul 11, 2026
- Updated
- Jul 11, 2026
- SHA
- ac674715
Base models
Model card
From the Hugging Face model README
Ortho Vision QLoRA — FINAL v4
Final v4 LoRA adapter for orthographic projection behavior.
Given an image of a unit-cube solid (unit grid + marked FRONT edge), the model outputs:
dims: ZxYxX
top:
...
front:
...
right:
...
Base model
unsloth/Qwen2.5-VL-3B-Instruct (4-bit QLoRA / Unsloth)
Training data
hiyasvyas/orthographic-views-vision-final-v4
- ~5400 train image–label pairs
- Held-out golden eval (100) excluded from train
Training setup
| Setting | Value |
|---|---|
| Method | QLoRA (Unsloth) |
| LoRA rank | 32 |
| LR | 2e-4 |
| Epochs | 2 |
| Batch | 2 × grad_accum 4 |
Eval (held-out golden, n=100)
| Metric | TUNED (final v4) |
|---|---|
| format_ok | 98% |
| dims_ok | 88% |
| top_ok | 41% |
| front_ok | 51% |
| right_ok | 49% |
| exact | 23% |
Base model exact ≈ 0% on the same set (prompt litmus fails). Exact wins byte-identical to gold; train/golden label leakage 2/100.
Load
from unsloth import FastVisionModel
model, tokenizer = FastVisionModel.from_pretrained(
"hiyasvyas/ortho-vision-qlora-final-v4", # or load base then attach adapter
load_in_4bit=True,
)
# If this repo is adapter-only, load base then:
# model.load_adapter("hiyasvyas/ortho-vision-qlora-final-v4")
Behavior Spec
Given an image of a unit-cube solid (grid + FRONT edge), output
dims: ZxYxXthen exactly three ASCII gridstop:/front:/right:(third-angle), only./#, sizes matching dims.
- Library
- peft
- License
- mit
- Created
- Jul 11, 2026
- Updated
- Jul 11, 2026