shiyunliu/IR_mllm

IR MLLM LoRA Adapters

This repository contains LoRA adapters trained for the IR_mllm geometry intermediate-language experiments.

Included adapters

• models/formalgeo_route_qwen17b_lora/: Qwen3-1.7B LoRA trained to map question + GDP-4B parse to theorem-route language using FormalGeo theorem sequences.
• models/gdp_to_plan_qwen17b_lora_8ep_final/: earlier Qwen3-1.7B LoRA trained to map question + GDP-4B parse to construction-plan language.

Main route-generator training summary

• Source: FormalGeo theorem sequences.
• GDP parsed images: 900.
• Split: 600 train / 100 validation / 200 downstream test.
• Final validation loss / perplexity: 0.13796 / 1.14792.

Downstream Geometry3K-150 result: image-only 50.67%, image + generated theorem route 54.67%.