MElHuseyni/SauerkrautLM-ColQwen3-2b-v0.10-tr

Turkish ColQwen3 - Full Fine-Tuned

This model is a fully fine-tuned version of VAGOsolutions/SauerkrautLM-ColQwen3-2b-v0.1 for Turkish document retrieval.

Training Details

• Training Type: Full Fine-Tuning (NOT LoRA/Adapter)
• Base Model: VAGOsolutions/SauerkrautLM-ColQwen3-2b-v0.1
• Training Datasets:
  • selimc/tr-textbook-ColPali - Turkish textbook pages
  • muhammetfatihaktug/bilim_teknik_mini_colpali - Bilim Teknik magazine pages
• Learning Rate: 2e-05
• Epochs: 1
• Effective Batch Size: 16

Usage

import torch
from sauerkrautlm_colpali.models import ColQwen3, ColQwen3Processor

# Load the full fine-tuned model directly (no adapter loading needed!)
model = ColQwen3.from_pretrained(
    "MElHuseyni/turkish-sauerkrautlm-colqwen3-full",
    torch_dtype=torch.bfloat16,
    device_map="cuda:0"
).eval()

processor = ColQwen3Processor.from_pretrained("MElHuseyni/turkish-sauerkrautlm-colqwen3-full")

# Process images and queries
images = [...]  # Your images
queries = ["Bu belgede ne anlatılıyor?"]

batch_images = processor.process_images(images)
batch_queries = processor.process_queries(queries)

# Get embeddings
with torch.no_grad():
    image_embeddings = model(**batch_images.to(model.device))
    query_embeddings = model(**batch_queries.to(model.device))

# Calculate similarity scores
scores = processor.score(query_embeddings, image_embeddings)

Differences from LoRA Version

Aspect	This Model	LoRA Version
Loading	Direct load	Requires base + adapter
Size	Full model (~4-8GB)	Small adapter (~50MB)
Performance	Potentially better	Good
Training	All weights updated	Only adapter weights

License

Apache 2.0