Blame for README.md

---
title: Minecraft Agent
emoji: ⛏️
colorFrom: green
colorTo: blue
sdk: static
pinned: false
models:
- EmberJin/Minecraft-Agent-Qwen3.5-9B-Stage3
---

# Minecraft Agent — Qwen3.5-9B (Stage3)

A vision-language agent that plays Minecraft from raw first-person pixels (640×360 @ 20 FPS) and emits structured text actions. Stage-3 SFT on 200k+ human gameplay trajectories (6M frames, 20 FPS) converted to a text-action format.

## Pages

- **[`index.html`](index.html)** — Model overview, evaluation table, training recipe, action space, and a real annotated trajectory from the training data.
- **[`cot.html`](cot.html)** — Demo dataset: 531 trajectories of decision-point Chain-of-Thought annotations (CoT pilot subset).

## CoT Demo Dataset

The CoT demo dataset (531 trajectories, 1,917 thoughts) is available under `data/`:

| File | Size | Contents |
|---|---|---|
| `data/thoughts.jsonl` | ~3 MB | Text-only export: instruction + step / thought / action |
| `data/annotated.pkl` | 450 MB | Full trajectories with image bytes (for continue-SFT training) |

Annotation pipeline (in `cot_annotation/`): decision-point detection (R1/R2/R3 rules, ~12.4% trigger rate) → VLM dual-channel prompting → step-alignment + format validation (≤45 words) → rendering. See the model repo's research notes for the design rationale.