patnir41/kaetram-qwen3.5-2b-opd-r2

Kaetram Qwen3.5-2B — OPD Round 2

A 2B agent that plays Kaetram by calling typed game tools. This is round 2 of on-policy distillation (OPD) — the round where the first weights-driven competence gain appears.

Core-3 = quest-stage progress summed across three archetype agents (10 stages each, 30 max) over the Foresting → Herbalist's Desperation → Rick's Roll quest chain.

Round 2 broke a wall the base and r1 never passed: all three archetype agents pass the Herbalist's Desperation stage-1 wall unseeded (base and r1 passed it 0/3; r3 goes on to finish the chain). The lift is attributable to the weights — round 2 changed only the policy, not the harness — confirming the on-policy distillation signal instilled new task competence, not just style.

Method

On-policy distillation with a reverse-KL advantage against a scaffolded 4B teacher, advantage = -(logp_student − logp_teacher), trained with PPO-clipped importance sampling (LoRA r=64, α=64, no rsLoRA, bf16, 1 epoch, advantage clamp ±3, early-turn step-weight 1.5). Round 2 fits a fresh LoRA on the merged r1 checkpoint and is trained on rollouts from the r1 policy plus a seeded "wall-state" collection. Full construction: `patnir41/kaetram-opd-2b`.

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

Files

• —root: merged bf16 weights (Qwen3_5ForConditionalGeneration) — load directly.
• —adapter/: the LoRA adapter alone (applies on top of the merged r1 checkpoint).

Text-only fine-tune of a multimodal-capable base; chat_template.jinja preserves <think> on every assistant turn.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained("patnir41/kaetram-qwen3.5-2b-opd-r2", torch_dtype="bfloat16", device_map="auto")
t = AutoTokenizer.from_pretrained("patnir41/kaetram-qwen3.5-2b-opd-r2")

Limitations

Single-task agent for the Kaetram Core-3 benchmark. Known failure modes: a malformed tool-call attractor (mitigated, not eliminated) and the "Rick's Roll" quest, which stays unsolved across the whole program because the same-family teacher cannot grade a skill it cannot itself perform.

License & credits

Apache-2.0, inheriting Qwen3.5-2B (© 2026 Alibaba Cloud). Game environment/data 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.

Citation

@misc{kaetram_opd_2b_r2_2026,
  title        = {Kaetram Qwen3.5-2B OPD (Round 2)},
  author       = {patnir41},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/patnir41/kaetram-qwen3.5-2b-opd-r2}}
}