GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Abdoul27
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
Abdoul27
Model Tree
Qwen/Qwen3-VL-32B-Instruct
Input Modalities
Output Modalities
Supported Functionality
Metric (competition metric): score = 1 − WER_w/24 − CER_w/110, with word and character edit distances weighted per line by √(reference length). Greedy decoding, base loaded in 4-bit NF4 (as trained), image scale 1.0.
| split | lines | score | WER_w | CER_w | exact lines |
|---|---|---|---|---|---|
| validation ("dev150" = dev250 minus original100) | 150 | 0.8799 | 1.9839 | 4.1210 | 31 |
| original100 | 100 | 0.8838 | 2.0590 | 3.3434 | 19 |
| validation, epoch 1 checkpoint | 150 | 0.8660 | 2.2021 | 4.6464 | 28 |
| split | lines | use |
|---|---|---|
| train3497 | 3,497 | training only |
| validation = dev250 minus original100 | 150 | checkpoint selection (epoch 1 vs epoch 2) |
| original100 | 100 | reported once; never used for any choice |
| audit350 | 350 | holdout |
| test | 1,374 | competition test lines (no labels) |
Tested with:
torch==2.4.1+cu124transformers==4.57.3peft==0.21.1bitsandbytes==0.50.2accelerate==1.15.0plus Pillow, huggingface_hub, optionally flash-attn. GPU: ~24 GB for the 4-bit base + adapter (the setting that was scored);
~70 GB for a bf16 base.
bash
pip install "transformers==4.57.3" "peft==0.21.1" "bitsandbytes==0.50.2" accelerate pillow huggingface_hubhuggingface-cli login # after your access request is approvedpython
from huggingface_hub import snapshot_downloadimport syspath = snapshot_download("Abdoul27/qwen3-vl-32b-barbados-lora")sys.path.insert(0, path)from barbados_inference import load, transcribemodel, processor = load(path) # downloads Qwen/Qwen3-VL-32B-Instruct (~65 GB) on first useprint(transcribe(model, processor, ["line_001.jpg", "line_002.jpg"], batch_size=4))What barbados_inference.py does (reproduce this exactly to get the scores above):
barbados_inference.py / barbados_inference.json),
then the image and the user prompt "Transcribe the handwritten line in this image.", through the model's chat template with the generation prompt.Qwen/Qwen3-VL-32B-Instruct in 4-bit NF4 (double quantisation, bf16 compute; vision tower and lm_head kept in bf16) + this adapter.Transcription: and surrounding quotes removed, spaces normalised.Transcription conventions: original spelling kept; abbreviations copied letter for letter; ^ marks raised letters (W^m, w^th,
Exec^rs); & kept; long s written as s; numerals as digits.
QLoRA on the 4-bit NF4 base: LoRA r=16, α=32, dropout 0.05 on the language-model attention and MLP projections (vision tower frozen), AdamW lr 0.0001, cosine schedule with 3 % warm-up, batch 1 × 16 gradient accumulation, 2 epochs, answer-only loss, mild augmentation (±10 % scale, brightness/contrast). Best checkpoint: epoch 2, selected on the validation split.
Apache-2.0, as the base model Qwen/Qwen3-VL-32B-Instruct.
pip install "transformers==4.57.3" "peft==0.21.1" "bitsandbytes==0.50.2" accelerate pillow huggingface_hubhuggingface-cli login # after your access request is approvedfrom huggingface_hub import snapshot_downloadimport syspath = snapshot_download("Abdoul27/qwen3-vl-32b-barbados-lora")sys.path.insert(0, path)from barbados_inference import load, transcribemodel, processor = load(path) # downloads Qwen/Qwen3-VL-32B-Instruct (~65 GB) on first useprint(transcribe(model, processor, ["line_001.jpg", "line_002.jpg"], batch_size=4))