---
datasets:
- lewiswatson/Frame2KG-YC2
pipeline_tag: image-text-to-text
tags:
- Frame2KG
---
# Frame2KG Adapter Archive

This repository contains archived LoRA checkpoints for Frame2KG.

If you want to use Frame2KG in code, use the canonical adapter repositories instead.  
This archive exists for reproducibility and research comparison.

---

## Canonical adapters

- lewiswatson/frame2kg-qwen2_5vl-3b-qkvo-lora
- lewiswatson/frame2kg-qwen2_5vl-3b-qkvo-gateupdown-lora
- lewiswatson/frame2kg-qwen2_5vl-7b-qkvo-lora
- lewiswatson/frame2kg-qwen2_5vl-7b-qkvo-gateupdown-lora

---

## Current contents

This archive currently contains LoRA checkpoints trained on:

- Qwen2.5-VL 3B
- Qwen2.5-VL 7B

Each top-level directory corresponds to a training variant.  
Each variant contains four shortlisted checkpoints:

- 1 (Global step 1000)
- 2 (Global step 2000)
- 3 (Best/lowest training loss)
- 4 (Final checkpoint)

Each checkpoint folder contains standard PEFT adapter weights.

---

## Chat template

The file `chat_template.jinja` is included for consistency with training and evaluation.

---

Additional model backends may be added in future.