A small vision-language model that captions short clips of a captive Octopus vulgaris ("Nity"), distilled from a large teacher (Qwen3-VL-235B) via QLoRA. Part of an automated pipeline that turns continuous aquarium footage into a behavioural time-series.

mlx_4bit/ — merged + 4-bit MLX build (~1.7 GB). Runs on Apple Silicon, ~3 s/caption, no GPU.
lora/ — the LoRA adapter (r16/α32) for the cross-platform path: base Qwen/Qwen3-VL-2B-Instruct + adapter.

Results (50 held-out val clips, vs. base model)

Metric	Base	This model
Embedding similarity to teacher	0.702	0.834
ROUGE-L	0.269	0.455

Usage (MLX, Apple Silicon)

python

from mlx_vlm import load, generate
model, proc = load("sidraj000/octopus-nity-caption-qwen3vl2b-mlx-4bit", adapter_path=None)  # points at mlx_4bit/

For CUDA/CPU: load base Qwen/Qwen3-VL-2B-Instruct + the lora/ adapter via PEFT.

Limitations

Trained on footage of a single animal from one facility over ~months; evaluated against a teacher model, not human-verified ground truth. Behaviour phrasing reflects the teacher's biases.

Ethics

Footage collected in a private aquarium under sub-threshold interactions (EC directive 2010/63/EU), FELASA-compliant husbandry. The model contains no raw footage.

sidraj000/octopus-nity-caption-qwen3vl2b-mlx-4bit

Octopus "Nity" Caption Student (Qwen3-VL-2B, distilled)

Contents

Results (50 held-out val clips, vs. base model)

Usage (MLX, Apple Silicon)

Limitations

Ethics

sidraj000/
octopus-nity-caption-qwen3vl2b-mlx-4bit