RESEARCH-EMPRM/emprm-v2-a2_support_qwen25_s0

EM-PRM v2 — a2_support_qwen25_s0

LoRA adapter for Qwen/Qwen2.5-VL-7B-Instruct from the paper EM-PRM: Evidence-Mediated Process Rewards for Robust Multimodal Reasoning (EM-PRM v2 experiment ladder).

A2 support checkpoint on Qwen2.5-VL-7B-Instruct (initialisation of the rung-6 arms).

Training

• LoRA rank 64, alpha 128, dropout 0.05, target modules k_proj, layers.0.mlp.down_proj, layers.0.mlp.gate_proj, layers.0.mlp.up_proj, layers.1.mlp.down_proj, layers.1.mlp.gate_proj, layers.1.mlp.up_proj, layers.10.mlp.down_proj, layers.10.mlp.gate_proj, layers.10.mlp.up_proj, layers.11.mlp.down_proj, layers.11.mlp.gate_proj, layers.11.mlp.up_proj, layers.12.mlp.down_proj, layers.12.mlp.gate_proj, layers.12.mlp.up_proj, layers.13.mlp.down_proj, layers.13.mlp.gate_proj, layers.13.mlp.up_proj, layers.14.mlp.down_proj, layers.14.mlp.gate_proj, layers.14.mlp.up_proj, layers.15.mlp.down_proj, layers.15.mlp.gate_proj, layers.15.mlp.up_proj, layers.16.mlp.down_proj, layers.16.mlp.gate_proj, layers.16.mlp.up_proj, layers.17.mlp.down_proj, layers.17.mlp.gate_proj, layers.17.mlp.up_proj, layers.18.mlp.down_proj, layers.18.mlp.gate_proj, layers.18.mlp.up_proj, layers.19.mlp.down_proj, layers.19.mlp.gate_proj, layers.19.mlp.up_proj, layers.2.mlp.down_proj, layers.2.mlp.gate_proj, layers.2.mlp.up_proj, layers.20.mlp.down_proj, layers.20.mlp.gate_proj, layers.20.mlp.up_proj, layers.21.mlp.down_proj, layers.21.mlp.gate_proj, layers.21.mlp.up_proj, layers.22.mlp.down_proj, layers.22.mlp.gate_proj, layers.22.mlp.up_proj, layers.23.mlp.down_proj, layers.23.mlp.gate_proj, layers.23.mlp.up_proj, layers.24.mlp.down_proj, layers.24.mlp.gate_proj, layers.24.mlp.up_proj, layers.25.mlp.down_proj, layers.25.mlp.gate_proj, layers.25.mlp.up_proj, layers.26.mlp.down_proj, layers.26.mlp.gate_proj, layers.26.mlp.up_proj, layers.27.mlp.down_proj, layers.27.mlp.gate_proj, layers.27.mlp.up_proj, layers.3.mlp.down_proj, layers.3.mlp.gate_proj, layers.3.mlp.up_proj, layers.4.mlp.down_proj, layers.4.mlp.gate_proj, layers.4.mlp.up_proj, layers.5.mlp.down_proj, layers.5.mlp.gate_proj, layers.5.mlp.up_proj, layers.6.mlp.down_proj, layers.6.mlp.gate_proj, layers.6.mlp.up_proj, layers.7.mlp.down_proj, layers.7.mlp.gate_proj, layers.7.mlp.up_proj, layers.8.mlp.down_proj, layers.8.mlp.gate_proj, layers.8.mlp.up_proj, layers.9.mlp.down_proj, layers.9.mlp.gate_proj, layers.9.mlp.up_proj, o_proj, q_proj, v_proj; vision tower frozen; bfloat16.
• Seed 0, learning rate 5e-05, effective batch 2×4, one epoch.
• Training data, pair sets and every gate artifact are in the mirror RESEARCH-EMPRM/emprm-v2 (dataset repo; results/runs_v2/train/a2_support_qwen25_s0/) and the paper bundle under backdata/.

Load

from transformers import AutoModelForImageTextToText, AutoProcessor
from peft import PeftModel
base = AutoModelForImageTextToText.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct", dtype="bfloat16", device_map="cuda")
model = PeftModel.from_pretrained(base, "RESEARCH-EMPRM/emprm-v2-a2_support_qwen25_s0")
processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct")

adapter_config.json records the local path the adapter was trained from; pass the base model explicitly as above. Scoring prompts (bank extraction, claim extraction, claim support, ranking) are the ones in work/scripts/eval_bon.py of the mirror.

Provenance

Trained in the EM-PRM v2 repository; every number quoted in the paper is traceable to planning/V2_PLAN.md and the generated tables in the mirror.