CoolFace
Modelpublic

RESEARCH-EMPRM/emprm-v2-stageB_rung5_noev_s2

sourceHugging Faceapache-2.0updated 17d agoView on Hugging Face
0likes18downloads
Model Card

EM-PRM v2 — stageB_rung5_noev_s2 — R+G arm (E2 ablation), seed 2

LoRA adapter for Qwen/Qwen3-VL-8B-Instruct (snapshot 0c351dd) from the EM-PRM v2 experiment ladder (EM-PRM: Evidence-Mediated Process Rewards for Robust Multimodal Reasoning). Trained 2026-09-10 at git commit 4438aea.

The rung-5 recipe with the evidence line removed from the ranker prompt at training and inference (GPRM_RANKER_NO_EVIDENCE=1): same 20,000 task records, the same 8,353 pairs rendered without the evidence line, same A2 initialisation, same schedule, pair_sees_image: true. Its product pass is the R+G cell of the E2 factorial (record gate x evidence-free ranker); its verify_only pass is the R cell.

Training

  • —LoRA rank 64, alpha 128, dropout 0.05, target modules downproj, gateproj, kproj, oproj, qproj, upproj, v_proj (174,587,904 trainable parameters); vision tower frozen; bfloat16.
  • —Seed 2, learning rate 5e-05, micro-batch 2 x gradaccum 4, 1 epoch, 2,500 optimiser steps over 20,000 task records; pair records 8,353, pair micro-batches 5,000, lambdapair 1.0, pairseesimage True.
  • —Initialised from a2_support_s0 (RESEARCH-EMPRM/emprm-v2-a2_support_s0). Task-data sha256 46e0cb91457a…, pair-data runs/v2/data/stageB_pairs_rung5_noev/pairs.jsonl.
  • —Wall time 5.5 h on one NVIDIA A100-PCIE-40GB; training-pair accuracy mean 0.8528, final 0.97.

Pre-registered gates and reads (development halves; test halves unread)

  • —Held-out relational FlipAcc (deployed): 0.0000 [0.0000, 0.0000] — gate lost, forced by the architecture; 0.0037 with the chart shown.
  • —Forced-evidence acceptance at 0.5: true 0.4900 / false 0.005 deployed; 0.925 / 0.005 with the chart shown.
  • —Chart-disjoint pair gain over v1: +0.0891 [0.0440, 0.1336] deployed; +0.1274 with the chart shown.
  • —Controlled pools (dev, Best-of-5, InternVL / Qwen): 0.3305 / 0.4670 (paired against the final arm: -0.5064 [-0.5405, -0.4723] / -0.3806 [-0.4147, -0.3486]).
  • —External dev halves (deployed): VisualProcessBench 0.4350, VLRMBench 0.3959, VL-RewardBench 0.5413, Multimodal RewardBench 0.5051.

Status

Not a deployment candidate (chart gates lost). Kept as the seed-2 replicate of the "E removed" ablation.

Where the artifacts are

  • —Result files, per-example dumps, config and prompts: dataset RESEARCH-EMPRM/emprm-sync-20260910 → results/**/runs/v2/train/stageB_rung5_noev_s2/, results/**/runs/v2/e2/stageB_rung5_noev_s2__*.json, configs/ablations/stageB_rung5_noev_s2.yaml, EXPERIMENT_REGISTRY.csv (rows tagged with this adapter), CURRENT.md and WRITER_SYNC_BUNDLE.md (what the deployed scorer computes; which arm is which).
  • —The frozen 2026-09-09 tree backup RESEARCH-EMPRM/emprm-v2 predates this arm and does not contain it.

Load

python
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_noev_s2")
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. The scorer (scoring.Scorer.score_grounded, family grounded, aggregation product) and its prompts are in code/ of the sync dataset.