yoonLM/mal2026-r0-ensemble-v1

MAL2026 R0 P1--4 prediction ensemble

This repository contains the custom LoRA adapters and three-axis regression heads used by the MAL2026 metric-first submission candidate. It does not contain competition train/validation rows, writing text, identifiers, predictions, generated row outputs, optimizer state, or credentials.

Architecture

1. skt/A.X-4.0-Light generates score-blind rationales with the pinned rank2_ax4_random1 adapter.
2. Qwen/Qwen3-Embedding-8B reads the prompt, essay, and three rationales.
3. Epoch 1--4 LoRA/head predictions are averaged uniformly as continuous values, clipped to [1, 5], and rounded half-up for official integer output.
4. The optional final DPO adapter explains the emitted scores.

Pinned upstream revisions and local artifact checksums are recorded in bundle_complete.json. The Docker submission embeds the upstream base model snapshots and performs no network download at container startup.

Development evidence

• continuous macro RMSE: 0.5582937519
• continuous macro Spearman: 0.6441959865
• integer macro RMSE: 0.6158981882
• integer macro Spearman: 0.5681974395

These are previously exposed 400-row validation results and must not be interpreted as an untouched hidden-test estimate.