This repository contains full merged weights: a fine-tune of Qwen/Qwen3.5-2B for interactive dialogue-game play, trained with parameter-efficient methods on colab-potsdam/playpen-data (train split only). The scaled LoRA adapter used to produce these merged weights is preserved under lora-adapter/ for provenance.

Weights are stored in fp32 deliberately. The fine-tune delta is small relative to bf16's resolution at these weight magnitudes; materializing the merge in bf16 measurably degrades static-benchmark behavior (statscore 45.06 → 43.15 in our tests). Load with the default torch_dtype="auto" and do not downcast the weights.

Public validation results (dataset revision 557d8caf):

Model	Clemscore	Statscore	Δ Clem	Δ Stat
Official `Qwen/Qwen3.5-2B` baseline	13.05	44.02	—	—
This model	50.43	44.90	+37.38	+0.88

Clemscore is a paired same-host mean of two runs (50.82 / 50.03). The model is positive on both public axes versus the official base, and its clemscore exceeds the official 9B (41.92) and 4B (39.70) baselines. Independent verification of this repository's merged weights, loaded exactly as the public evaluator does (plain transformers, torch_dtype="auto"): clem 48.70 / stat 45.06 — both within suite variance of the headline runs, statscore still positive vs the official base. (The adapter path under lora-adapter/ verifies at clem 51.22 / stat 45.36.)

How to load

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "chnln/Qwen3.5-2B-playpen-playornotplay",
    trust_remote_code=True,  # weights are fp32 on purpose — do not downcast (see above)
)
tokenizer = AutoTokenizer.from_pretrained("chnln/Qwen3.5-2B-playpen-playornotplay")

All evaluations used the bundled chat template with enable_thinking=False.

Evaluation protocol

Interactive suite: playpen eval <model> --suite clem; static suite: playpen eval <model> --suite static
Public validation basis: colab-potsdam/playpen-data revision 557d8cafd1bc4557bc22803a6d4502ef53fb664c
model.safetensors merged from base + adapter; adapter SHA-256: fbcfa37fd1d30e45b85ae4b8e44b59bb1d1792b187b280f487629f12026c0d2e

Training

Trained by team playornotplay on the Playpen interaction/instance data (train split; no validation or private test data used). Summary training card:

Methodology: multi-stage LoRA fine-tuning (r=16, α=32, all-linear) — broad SFT on successful game episodes, then game-targeted SFT, then preference optimization (DPO) stages targeting diagnosed decision-level failure modes, followed by a training-free LoRA-delta scaling step (s=0.85) that recovers static-benchmark performance while preserving interactive gains.
Data usage breakdown: all data derives from the official colab-potsdam/playpen-data train split. SFT stages: success-only interaction transcripts (per-game balanced, ≤700 episodes/game; a weak-game subset filtered by source model). DPO stages: self-constructed preference pairs — ~3.8k synthetic turn-local pairs targeting mechanical failure modes, plus on-policy pairs collected by branching the current model during game play. No external datasets, no validation/test data.
Key hyperparameters: LoRA r=16 / α=32 / dropout 0.05; SFT LR 5e-5, max_len 2048; DPO LR 5e-6, β=0.1–0.2, max_len 1024; final delta scale s=0.85.
Compute budget: ~5 A100-hours total for the shipped lineage (single GPU, measured per job from scheduler logs).
Design decisions for reproducibility: dataset revision pinned and recorded per run (557d8caf basis); evaluation-harness commit pinned; per-run metadata (code SHA, environment, dataset revision); artifact SHA-256 checksums verified train-host → upload → Hub; merged weights stored fp32 (see warning above) and re-verified end-to-end from a clean Hub pull.

The full training card — per-stage configurations, data-construction procedures, per-stage results, and the complete failure-mode analysis — will be released with the workshop paper.

Limitations

Specialized for Playpen dialogue games; not intended as a general-purpose chat model.
Public clem validation is small (~42 instances) and noise-dominated (~3 pts run-to-run); statscore (430 instances) is more stable.
Scores depend on the dataset revision used by the public evaluator.

chnln/Qwen3.5-2B-playpen-playornotplay

Qwen3.5-2B Playpen Agent — team playornotplay

How to load

Evaluation protocol

Training

Limitations

chnln/
Qwen3.5-2B-playpen-playornotplay

Qwen3.5-2B Playpen Agent — team `playornotplay`