RESEARCH-EMPRM/emprm-v2-stageB_rung5_s1
06
EM-PRM v2 — stageB_rung5_s1
LoRA adapter for Qwen/Qwen3-VL-8B-Instruct from the paper EM-PRM: Evidence-Mediated Process Rewards for Robust Multimodal Reasoning (EM-PRM v2 experiment ladder).
EM-PRM rung 5, primary backbone, seed 1 (same initialisation and data as seed 0).
Training
- LoRA rank 64, alpha 128, dropout 0.05, target modules downproj, gateproj, kproj, oproj, qproj, upproj, v_proj; vision tower frozen; bfloat16.
- Seed 1, 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/stageB_rung5_s1/) and the paper bundle underbackdata/.
Pre-registered gates (development half)
- Held-out relational FlipAcc (operation subset): 0.7638
- Forced-evidence acceptance at 0.5: false 0.01, true 0.965
- Chart-disjoint pair gain against the v1 head: +0.1026 (95% interval [0.0614, 0.1443])
Load
from transformers import AutoModelForImageTextToText, AutoProcessor
from peft import PeftModel
base = AutoModelForImageTextToText.from_pretrained("Qwen/Qwen3-VL-8B-Instruct", dtype="bfloat16", device_map="cuda")
model = PeftModel.from_pretrained(base, "RESEARCH-EMPRM/emprm-v2-stageB_rung5_s1")
processor = AutoProcessor.from_pretrained("Qwen/Qwen3-VL-8B-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.
