AdVSim2Real Stage 2: With Stage 1
Stage 2 adversary/executor co-evolution with 150 proposals per iteration, 1 adversary epoch, 2 executor epochs, 3 iterations.
The curriculum stays frozen at Stage 1 curr_v3. The first executor starts from
Stage 1 exec_v3. Both are published in Sarim-Hash/advsim2real-stage1-curr1epoch-exec2epochs-iter3-nprop150
at commit f9c3ef2a5c8364fac8312f192e9f803cff17d264. The first adversary starts from the base model.
Later adversaries continue the preceding Stage 2 adversary adapter; later executors continue the preceding Stage 2 executor adapter. The curriculum remains frozen throughout. The World model is WebWorld-14B. These are experimental training artifacts; publication does not assert an independently evaluated score.
Adapter folders appear only after their full training schedule completes and
their root checkpoint passes completion and byte-hash checks: adv_v1, exec_v1,
adv_v2, exec_v2, adv_v3, exec_v3. An absent folder is still pending.
This repository can exist before the first adapter finishes.
Load Qwen/Qwen3.5-4B at revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a, then apply an available
LoRA adapter using PEFT:
from peft import PeftModel
model = PeftModel.from_pretrained(base_model, "Sarim-Hash/advsim2real-stage2-with-stage1-adv1epoch-exec2epochs-iter3-nprop150", subfolder="exec_v1")
Every adapter contains exactly adapter_model.safetensors, adapter_config.json,
and training.json. Weight bytes match the finalized training checkpoint; the
configuration identifies the public upstream base instead of a local machine path.
Training metadata reports attempted steps, actual optimizer updates and available
policy-signal counters. Completed schedules can include skipped optimizer updates.
Other fixed settings: 4 clean rollouts per task, adversary groups of 6, executor groups of 4, 2 prompts/tasks per batch, and maximum executor training length 4096. Per-adapter metadata records the remaining numeric training settings and lineage.
- Downloads last month
- -