EnvScaler Qwen3-8B GRPO Step 63 (2026-05-15)
This repository is a recoverable Hugging Face-format export of the terminal checkpoint from an early EnvScaler non-conversational 32-GPU GRPO run.
Provenance
- Internal run:
envscaler_non_conv_rl_grpo_32gpu_20260515_225741 - Base checkpoint recorded by the run:
EnvScaler-Qwen3-8B-SFT - Selected checkpoint: step 63
- Architecture: Qwen3 causal LM, approximately 8.19 billion parameters
- Original format: two-way tensor-parallel Megatron checkpoint
- Export format: Hugging Face safetensors
- Training period: 2026-05-15 to 2026-05-16
The run completed normally after 64 recorded steps (0 through 63). It was an early experiment and was not referenced by the formal BFCL, VitaBench, or Tau2 evaluation workflows found during the archive audit. It should therefore not be interpreted as a benchmark-selected release.
Training Metrics
training_metrics.jsonl contains only numeric per-step metrics extracted from
the driver log. Prompts, tool arguments, environment content, and credential
fields were deliberately excluded. training_summary.json provides the main
aggregate values:
- Mean rollout score: 0.704125
- Peak rollout score: 0.851259 at step 41
- Final rollout score: 0.660812 at step 63
These rollout scores are online training signals, not held-out benchmark results. A higher single-step value does not establish that the corresponding checkpoint is better than the terminal checkpoint.
Conversion Validation
The converter reconstructed and validated all 399 expected Qwen3 parameters
from the two Megatron tensor-parallel model shards. It checked the complete key
set, tensor shapes, and dtypes before writing the safetensors files. The result
was then loaded successfully with AutoModelForCausalLM.from_pretrained and
reported 8,190,735,360 parameters. The tokenizer was also loaded successfully.
Optimizer state, RNG state, raw driver logs, and the other intermediate Megatron checkpoints are intentionally not included.