GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
hackiit-neeraj
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
hackiit-neeraj
Model Tree
Qwen/Qwen2.5-VL-3B-Instruct
Input Modalities
Output Modalities
Supported Functionality
python
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessorfrom peft import PeftModelbase_id = "Qwen/Qwen2.5-VL-3B-Instruct"model = Qwen2_5_VLForConditionalGeneration.from_pretrained(base_id, torch_dtype="auto", device_map="auto")model = PeftModel.from_pretrained(model, "hackiit-neeraj/qwen25vl-ahc-lora-ckpt400")processor = AutoProcessor.from_pretrained(base_id)
Trained against unsloth/qwen2.5-vl-3b-instruct-unsloth-bnb-4bit. The
official Qwen/Qwen2.5-VL-3B-Instruct weights are the same architecture
and load this adapter.
| Method | Unsloth QLoRA, 4-bit base, PEFT LoRA |
| Rank / alpha / dropout | r=16, alpha=16, dropout=0 |
| Targets | language-model attention + MLP projections |
| Hardware | Kaggle Tesla T4 |
| This checkpoint | warm-start from checkpoint-500, seed 1234, lr 3e-5, 400 steps, 0 non-finite losses |
| Eval | AHC public test, adaptive multi-frame, min_frames=2 → macro-F1 0.437 |
Do not load optimizer.pt / trainer state — those were stripped on purpose.
Offline / gated Stage-3 labelling of drone frames. A still frame cannot see motion, so this adapter should describe what is visible (and name an appearance class). Tracker / motion rules should own stop / stall / wrong-way decisions.
Adapter is Apache-2.0, same as the Qwen2.5-VL base. You need the base model weights separately.