Evidence selection:
- input: all slide grids + question
- target: gold evidence page numbers
Answer generation:
- input: gold evidence slide grids + question
- target: short answer

Training Setup

Framework: Unsloth + PEFT LoRA

Approximate config:

text

model: LiquidAI/LFM2.5-VL-1.6B
LoRA rank: 16
LoRA alpha: 16
vision layers: frozen
language layers: LoRA-tuned
attention modules: LoRA-tuned
MLP modules: LoRA-tuned
learning rate: 1e-4
max sequence length: 4096
precision: fp16/bf16 depending on hardware

Usage

python

import torch
from peft import PeftModel
from transformers import AutoModelForImageTextToText, AutoProcessor

base_model = "LiquidAI/LFM2.5-VL-1.6B"
adapter = "saad1926q/lfm2.5-vl-slidevqa-lora"

processor = AutoProcessor.from_pretrained(base_model, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    base_model,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
).to("cuda")
model = PeftModel.from_pretrained(model, adapter).eval()

Notes

This is a task-specific LoRA adapter, not a standalone model. Load it on top of LiquidAI/LFM2.5-VL-1.6B.

The adapter is intended for research on multi-slide document VQA and SlideVQA-style tasks.