---
title: MR V3 SFT Results Viewer
emoji: 📐
colorFrom: indigo
colorTo: green
sdk: static
app_file: index.html
pinned: false
license: mit
short_description: Browse Midtrain V3 SFT results + judged answers
---

# MR V3 SFT Results Viewer

A static viewer for **Midtrain V3** SFT *results* — one materialized trajectory
per problem, assembled from a results parquet (e.g.
[`HerrHruby/soda2026-mr-v3sft-results`](https://huggingface.co/datasets/HerrHruby/soda2026-mr-v3sft-results)).

For each problem it shows:

- the **problem** with its metadata (paper, prompt type, difficulty, target
  level, topic tags),
- the **layers of exploration** — each executed direction with its *found*,
  *rationale*, and *core result*,
- the synthesized **final answer**,
- the **judge verdict** (score, normalized %, explanation), and
- the **grading rubric** and **reference answer** it was judged against.

Use the sidebar to filter by prompt type / difficulty, sort by score, and search
by title / paper / topic; the toggles in the top bar control which views are shown.

This Space bundles the pre-built per-problem index (`data/`) so it is fully
self-contained — no external data fetch at runtime. Math renders with KaTeX and
prose with marked, both vendored locally under `vendor/`.

To regenerate or extend the index, see `build_index.py` in the source repo.