Downloads · 30 days
7
23% of all-time downloads
amanupg/wearable-ai-2026-egoconv-qwen3vl32b-qlora-r32
wearable-ai-2026-egoconv-qwen3vl32b-qlora-r32 is a image-text-to-text model from amanupg. 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 peft. The card lists the license as apache-2.0.
Runner-up entry (2nd of 7) in the EgoConv Large (2B+) sub-track of the Wearable AI Grand Challenge at ECCV 2026, hosted by Meta.
Downloads · 30 days
7
23% of all-time downloads
All-time downloads
30
Public
Repo size
1.1 GB
Likes
0
Public
Click a slice to open those files.
.safetensors1.1 GB · 99%
From the Hugging Face model README
Runner-up entry (2nd of 7) in the EgoConv Large (2B+) sub-track of the Wearable AI Grand Challenge at ECCV 2026, hosted by Meta.
Team SoloLevelling (solo entry), Aman Upganlawar, Columbia University.
| Result | 2nd of 7, runner-up |
| LLM-as-Judge (official, private test split) | 0.4816 |
| BLEU (official) | 0.1256, highest in the division |
| Total parameters served | 33.36 B (base; adapter merged at load) |
| Adapter | 896 tensors, 1.07 GB |
EgoConv is streaming multi-turn conversational QA over egocentric video. At each turn the
model sees the video up to that timestamp plus the dialogue so far and generates a free-form
answer, scored 1.0 / 0.5 / 0.0 per turn by Llama-4-Maverick-17B-128E-Instruct-FP8.
base Qwen3-VL-32B-Instruct
quantisation 4-bit NF4 (QLoRA), training-time only
lora_r 32
lora_alpha 64
lora_dropout 0.05
use_rslora true
target_modules q_proj k_proj v_proj o_proj gate_proj up_proj down_proj
epochs 1
frames 6 recent + 2 history, 896 px long side
One epoch, not two. A second epoch scored 0.4749 against this run's 0.5649 on our validation split, but that run was interrupted at 79% of a OneCycle schedule, so we report it as inconclusive rather than as evidence against a second epoch.
Quantisation was a training-time memory measure only. Merge the adapter and serve bf16.
The adapter was trained on 8 frames and served on 16. Measured with the official judge on a shared subset (150 videos, 1,014 turns, same adapter, only the serving budget differing):
| Serving frame budget | LLM-as-Judge |
|---|---|
| 6 recent + 2 history (as trained) | 0.4048 |
| 10 recent + 6 history (as served) | 0.5745 |
Removing the train/inference mismatch costs 17 points. At 8 frames the model drifts back into generic scene captioning instead of answering the question. Serve this adapter with the larger frame budget.
from peft import PeftModel
from transformers import AutoModelForImageTextToText, AutoProcessor
base = AutoModelForImageTextToText.from_pretrained(
"Qwen/Qwen3-VL-32B-Instruct", dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(
base, "amanupg/wearable-ai-2026-egoconv-qwen3vl32b-qlora-r32").merge_and_unload()
processor = AutoProcessor.from_pretrained("Qwen/Qwen3-VL-32B-Instruct")
github.com/amanupg/wearable-ai-2026-egoconv carries the training and evaluation code, the container recipe, raw logs, the submitted prediction file and a findings document covering the negative results and two retracted conclusions.
Apache-2.0, as required by the challenge prize terms (open weights plus a full source release under an OSI-approved permissive licence).