README.md

---
title: SFT v4 Dataset Viewer
emoji: 🧭
colorFrom: indigo
colorTo: purple
sdk: static
app_file: index.html
pinned: false
license: mit
short_description: Midtrain v4 SFT traces with repaired MR reasoning
---

# SFT v4 Dataset Viewer

Thirty meta-reasoning trajectories from the midtrain **v4** corpus, sampled
across the full range of depth (2–11 layers) and judge score.

Each trajectory shows, layer by layer:

- the **MR reasoning** — the assistant target SFT trains on, with the
  directions it commits to;
- the **executions** that pursued those directions, and their reports;
- the **summaries retained into the frontier**, which is what the next layer
  plans over;
- the **termination decision** and **final answer**, with the judge's verdict.

## About the MR reasoning shown here

Deliberation originally visited directions in F2 *rank* order. Rank encodes the
accept/reject decision, so the narrative came out sorted by disposition: every
kept direction before every discarded one in 96% of layers, against 40% in the
order the planner actually proposed them. Training on that teaches position
rather than judgement.

The generator now follows F1's proposal order, and the affected records were
regenerated. **This viewer shows only the repaired version** — trajectories
whose MR had not been regenerated are excluded rather than displayed with the
original ordering.