BeimingJingli/smart-fridge-qwen25vl-gguf

Smart Fridge Qwen2.5-VL 3B GGUF

Improved using Qwen / 基于 Qwen 微调。

这是面向智能冰箱场景微调并转换为 GGUF 的 Qwen2.5-VL 3B 模型。它用于接收 YOLO 预识别后的食物候选裁剪图，识别食物名称、类别和可见状态，并生成结构化的储存建议候选。

This is a Qwen2.5-VL 3B derivative fine-tuned for smart-fridge food-crop analysis and converted to GGUF for llama.cpp inference.

Files

File integrity hashes are provided in SHA256SUMS.

Intended Use

• —Smart-fridge food name and category recognition from cropped images.
• —Visible condition assessment, such as packaging state, color, shape and visible damage.
• —Structured observations for a downstream inventory database and rule engine.
• —Research, education and non-commercial evaluation only, as required by the base-model license.

The deployed project uses YOLO for fast candidate detection and change tracking. This VLM performs the main semantic analysis. Database history, storage duration and sensor readings are fused by downstream application logic rather than written directly by the model process.

llama.cpp Usage

Download both GGUF files:

hf download BeimingJingli/smart-fridge-qwen25vl-gguf \
  --local-dir smart-fridge-qwen25vl-gguf

Conservative CPU-only configuration used on a 4 GB RK3399 device:

llama-server \
  -m smart-fridge-qwen25vl-gguf/smart-fridge-qwen25vl-merged-Q4_K_M.gguf \
  --mmproj smart-fridge-qwen25vl-gguf/mmproj-smart-fridge-qwen25vl-Q8_0.gguf \
  --ctx-size 2048 \
  --parallel 1 \
  --threads 4 \
  --no-mmproj-offload \
  --image-min-tokens 64 \
  --image-max-tokens 64 \
  --jinja \
  --host 0.0.0.0 \
  --port 8080

The GGUF language-model metadata records a 128K architectural context length. The 2048-token setting above is an edge-device memory tradeoff, not the model maximum. For detailed localization or full-frame analysis, increase image tokens and context only after checking available memory.

The server exposes the OpenAI-compatible /v1/chat/completions endpoint. For this project, the prompt requests JSON fields including food_name, category, composition, freshness, freshness_score, visible_state, storage_advice, risk_level, confidence and notes.

Validation Status

Project integration checks completed with llama.cpp include:

• —Loading both GGUF components and reporting multimodal capability.
• —Encoding a real image and producing Chinese structured output.
• —Completing one cropped-food pipeline case through VLM normalization and SQLite ingestion.

These checks demonstrate integration compatibility only. They are not a standardized accuracy or food-safety benchmark. The training recipe, training dataset card and aggregate evaluation results are not included in this upload.

Limitations and Safety

• —Output is a visual estimate and can be wrong, especially under poor lighting, occlusion, unusual packaging or low-resolution crops.
• —The model cannot determine an expiry date, internal spoilage, contamination, allergens or food safety from an image alone.
• —Do not use its output as medical, regulatory or food-safety advice. Check labels, storage records, smell and appropriate professional guidance.
• —Sensor readings and inventory history must be validated by downstream logic; stale or estimated readings should not become definitive conclusions.
• —Quantization can reduce quality relative to the original precision model.

Base Model and License

Base model: Qwen/Qwen2.5-VL-3B-Instruct.

The base model is distributed under the Qwen Research License. It permits research and evaluation use but restricts use to non-commercial purposes unless a separate commercial license is obtained. Redistribution requires retaining the license and attribution notice. See LICENSE and NOTICE in this repository for the applicable terms.