old_poker

Legacy QwenPoker checkpoints and artifacts

This private repository preserves the locally available checkpoints and experiment outputs from earlier QwenPoker work. The contents are historical evidence. They do not establish playing strength against independent opponents.

Checkpoints

The RL adapter metadata records a local base path. Resolve the exact compatible base model and tokenizer before loading it outside the original environment. A LoRA adapter is not a standalone base model.

Earlier intermediate SFT and RL checkpoints also exist locally but are not included in this initial transfer. The table above is the exact checkpoint selection published here.

Artifacts

• outputs/ contains available training logs, configuration snapshots, evaluation tables, and reports. Four historical files with untranslated text or presentation content were excluded from the English release.
• artifacts/platform_evidence/ contains Phase 0 research, engine validation, and implementation evidence produced during later platform work. These are not completed model training weights or independently verified real hand histories.

The original raw OCR corpus and source videos are not included. Some historical records have incomplete experiment provenance. Do not treat a checkpoint directory name or a self-play metric as an independently validated model ranking.

The current platform code is maintained separately in runpodai09/poker_platform. New Qwen3.5-4B model weights are not part of this legacy repository because no completed local training checkpoint was identified for that model.