advsim2real-stage1-curr1epoch-exec2epochs-iter3-nprop150

AdVSim2Real Stage 1

Stage 1 curriculum/executor co-evolution with Qwen3.5-4B: 150 proposals, 1 curriculum epoch, 2 executor epochs, 3 iterations.

This run continues Stage 1 from an earlier handoff. The repository includes the handoff adapters exec_v1, curr_v1, and curr_v2, verified against the handoff hashes, followed by newly completed exec_v2, curr_v3, and exec_v3. A future folder is absent until its completion marker and adapter hashes pass validation. The continuation finishes executor E2, then curriculum C3 and E3.

These are LoRA adapters. Load the original Qwen/Qwen3.5-4B checkpoint (revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a), then select an available adapter folder:

from peft import PeftModel
model = PeftModel.from_pretrained(
    base_model, "Sarim-Hash/advsim2real-stage1-curr1epoch-exec2epochs-iter3-nprop150", subfolder="exec_v2"
)

Each folder includes adapter_model.safetensors, adapter_config.json, and training.json with completion provenance and SHA256 hashes. Weight bytes match the completed training checkpoint. The public configuration identifies the upstream base model instead of the training machine's local path.

Additional settings: 6 rollouts per task, 2 prompts per optimizer batch, maximum training sequence length 4096, learning rate 1e-5, KL coefficient 0.01. The world model is WebWorld-14B; trajectory judging uses DeepInfra.

Training trajectories, rewards, and logs

The available Stage 1 continuation logs are now published under training_logs/: 1,788 executor v2 trajectories, 1,800 executor v3 trajectories, 150 curriculum v3 task/reward records, training schedules and step logs, and separate task-assessment rollouts. See the log README for schemas, checkpoint provenance, and the missing original-machine logs for exec_v1, curr_v1, and curr_v2.