ConceptFormer-Phi3V

ConceptFormer-Phi3V

PEFT/LoRA checkpoint for ConceptFormer based on Tevatron/dse-phi3-docmatix-v1, initialized from the VDocRetriever Phi-3-Vision adapter. This repository contains the trained adapter and ConceptFormer sidecar, not a duplicate of the approximately 7.8 GB base model. Loaders resolve the base model from the base_model metadata and merge this adapter at load time.

Paper can be seen in Arxiv

Configuration

• Latent concept token: <|lcon|>
• Dynamic latent concept length from grounded regions
• Mean latent pooling
• Forward ranking-distribution KL, weight 0.2
• Three epochs, bfloat16, LoRA rank 8 / alpha 64 / dropout 0.1

Loading

import torch
from conceptformer.retriever.modeling import ConceptFormerRetriever

model = ConceptFormerRetriever.load(
    "Tevatron/dse-phi3-docmatix-v1",
    lora_name_or_path="hmhm1229/ConceptFormer-Phi3V",
    pooling="eos",
    normalize=True,
    dtype=torch.bfloat16,
    device_map="auto",
)
model.eval()

The wrapper loads the adapter tokenizer before PEFT, resizes the base embedding table for <|lcon|>, and merges the adapter. The base model is downloaded separately.

Results

Use this adapter with the ConceptFormer code repository. conceptformer_state.pt stores the latent projection used by the training objective.

Limitations

Performance depends on document rendering, image resizing, and corpus version. This model is intended for retrieval research and should not be used as a factual QA system without downstream validation.