A 2B agent that plays Kaetram (a 2D pixel MMORPG) by calling typed game tools. This is round 1 of an on-policy distillation (OPD) program that lifts a base Qwen3.5-2B toward a stronger 4B teacher on a 3-quest benchmark ("Core 3", scored out of 30 stages).

Round 1 transferred the teacher's style (turn economy, error rates, re-grounding habits) but not yet new task competence — the Core-3 score held at 12/30 while per-turn behavior measurably moved toward the teacher. The competence gains arrive in r2 and r3.

On-policy distillation: the student plays the game, and each emitted action token is scored with a reverse-KL advantage against the teacher, advantage = -(logp_student − logp_teacher). Training is PPO-clipped importance-sampling on those advantages (LoRA r=64, α=64, no rsLoRA, 7 projection modules, bf16, 1 epoch, advantage clamp ±3, early-turn step-weight 1.5). Round 1 initializes a fresh LoRA on base Qwen3.5-2B. Full construction is in the patnir41/kaetram-opd-2b dataset card.

Chain: base Qwen3.5-2B → r1 → (merge) → r2 → (merge) → r3.

This is a text-only fine-tune; the base architecture is multimodal-capable but no vision/audio path is trained or used. The included chat_template.jinja preserves <think> reasoning on every assistant turn.

The model emits typed tool calls (observe, navigate, attack, gather, query_quest, …) and expects the Kaetram MCP tool harness; outside that harness it generates the same tool-call syntax as plain text.

Trained for one narrow task (the Kaetram Core-3 benchmark) — not a general assistant. Inherits the round's known failure modes (occasional malformed tool-call syntax; the "Rick's Roll" quest stays unsolved across the whole program).

Apache-2.0, inheriting Qwen3.5-2B (© 2026 Alibaba Cloud). Game environment and embedded game data (coordinates, NPC/mob/quest names) are from Kaetram-Open (MPL-2.0). See NOTICE. All training data was generated by Qwen self-play — no third-party proprietary model outputs were used.