steven0226/qwen3vl-8b-chartqa-merged-16bit

Qwen3-VL-8B ChartQA — Merged 16-bit

Full merged checkpoint produced by merging `steven0226/qwen3vl-8b-chartqa-lora` into its Qwen3-VL-8B base. This is the quality reference used before AWQ quantization and the 16-bit serving baseline.

Training summary

• —15,000 shuffled ChartQA training examples
• —1 epoch; maximum sequence length 2,048
• —LoRA rank 16, alpha 16, dropout 0
• —Vision, language, attention, and MLP modules enabled for LoRA adaptation
• —Effective A100 batch size 16
• —8-bit AdamW; peak learning rate 2e-4; linear schedule
• —Runtime 3,579 seconds; final training loss 0.5907

Evaluation

Complete ChartQA test set, isolated vLLM evaluation, relaxed accuracy with 5% numeric tolerance:

Evaluation revision: 519060ef43df3261e0512e5ae4c82a4d4e675f32.

The downstream AWQ checkpoint scored 85.52%, a -0.72 pp change that passed the predefined -2 pp quality gate.

Serving reference

The formal A100 benchmark used this checkpoint as the 16-bit baseline. Weight files total 17.53 GB.

Complete comparison and reproducibility artifacts are in the AWQ repo `bench/`.

Usage

This is a standard merged Qwen3-VL checkpoint. It can be loaded with a Qwen3-VL-compatible Transformers/vLLM environment. The verified serving stack was vLLM 0.25.1+cu129 with torch 2.11.0+cu129 on Linux A100.

from vllm import LLM

llm = LLM(
    model="steven0226/qwen3vl-8b-chartqa-merged-16bit",
    revision="519060ef43df3261e0512e5ae4c82a4d4e675f32",
    max_model_len=4096,
    limit_mm_per_prompt={"image": 1, "video": 0},
)

For a complete image/chat example, use the AWQ model card and replace the model ID/revision.

Limitations

• —Evaluated primarily on English ChartQA short answers.
• —Not validated for general OCR documents, out-of-domain charts, or safety-critical use.
• —This checkpoint is the larger quality reference; the AWQ checkpoint is recommended for serving when its -0.72 pp quality tradeoff is acceptable.

中文摘要

這是將 ChartQA LoRA adapter 合併回 Qwen3-VL-8B 後的 16-bit 完整模型，也是 AWQ 量化前的品質與 serving baseline。完整 2,500 題 vLLM 評估為 86.24%；權重檔 17.53 GB。若部署可接受 0.72 個百分點的下降，建議改用 7.55 GB 的 AWQ 版本。

Source repository

程式碼、訓練與評估流程、測試與完整證據都在 GitHub：<https://github.com/kuotunyu/qwen3-vl-chartqa>。GitHub kuotunyu 與 Hugging Face steven0226 為同一人。

Source code, the training/evaluation pipeline, tests and the full evidence trail live at <https://github.com/kuotunyu/qwen3-vl-chartqa>. GitHub kuotunyu and Hugging Face steven0226 are the same author.