wiki-screenshot-text-baseline-lora

Wiki Screenshot Text Baseline LoRA

PEFT LoRA adapter from step 100 of a text-only contrastive baseline based on Qwen/Qwen3-VL-Embedding-2B. It was trained on Chrisyichuan/text-qa-pair with two mined hard negatives per query.

This checkpoint is an experimental fairness baseline for PixelRAG. The first 100 optimization steps used text query-passage pairs only; no screenshot training steps were performed.

Training configuration

• Steps at this checkpoint: 100 / 350
• Batch size: 64
• GradCache chunk: 4
• Hard negatives per query: 2
• Learning rate: 7e-6
• Warmup: 20 steps
• Scheduler: cosine
• LoRA rank / alpha: 32 / 32

Evaluation

The repository's existing evaluation loop uses the visual eval_hn.jsonl and miniv8 screenshot corpus, so these numbers are cross-modal diagnostics rather than a held-out text-corpus score.

A separate held-out text test split was not evaluated for this run.

Loading

from transformers import AutoModel
from peft import PeftModel

base = AutoModel.from_pretrained(
    "Qwen/Qwen3-VL-Embedding-2B",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(
    base,
    "Yichuanwwww/wiki-screenshot-text-baseline-lora",
)

Files

• adapter_model.safetensors: LoRA weights
• adapter_config.json: PEFT configuration
• SHA-256: a75a3f20ed8651ee3ebc45db3b213e808f4eb991b00fd9640c2fef515e65611e

The 4.46 GB optimizer/training state is intentionally excluded; these files are sufficient for inference and evaluation.