EnvACE 2.0 Non-Conversation GRPO 32-GPU (2026-05-20, checkpoint 200)
This repository stores the merged Hugging Face inference checkpoint produced by the aligned EnvScaler non-conversation GRPO run envscaler_non_conv_rl_grpo_32gpu_aligned_20260520_055146 at training step 200.
Contents
- Architecture:
Qwen3ForCausalLM(Qwen3-8B class) - Four
safetensorsmodel shards - Model index, configuration, tokenizer, generation configuration, and chat template
- 14 model/runtime files in total, excluding repository metadata and this README
Backup and recovery scope
The files were copied from the merged local export rl_runs/hf_models/envscaler_non_conv_rl_grpo_32gpu_aligned_20260520_055146_ckpt200.
On 2026-08-31, all 14 local files were verified byte-for-byte against this repository. LFS objects were checked using SHA256 and regular Git objects using Git blob hashes; no mismatch was found.
The matching logs and rollout dumps are stored separately in the dataset repository xuzishan/envace2.0-envscaler-grpo-32gpu-rollouts-20260520.
This repository preserves the deployable merged model. It does not preserve the original distributed training checkpoint, optimizer state, scheduler state, or dataloader state. It can be used for inference and evaluation, but it cannot reproduce an exact training resume from step 200.