ASIL Qwen3.5-2B RL
This repository contains the ASIL v0.1.0 paper release checkpoint for ASIL Qwen3.5-2B RL.
This model is presented in the paper ASIL: Replacing Screenshot-and-Click with Structured State and Semantic Actions.
- Release:
v0.1.0 - Selected checkpoint:
global_step_8_actor_hf - Source path:
/public/LLM_model_dataset/rl_temp/asil_operational_benchmark_a100_20260514_144617/results/rl/qwen35_2b_agentic_a800/qwen35_2b_rl_vllm_round3_final_small4_306039_20260519_114747/checkpoints/global_step_8_actor_hf - Base/init checkpoint:
/public/home/sjtu_normal/users/xierui/asil_sft_rl_a100_20260513_173133/results/sft_train/qwen35_2b_sft_v0_continue3_20260514_115443/checkpoints/global_step_27 - Training data:
rl_learnable_v4_320_80; 320 train / 80 valid task prompts - Prepared at:
2026-07-30T18:02:32+08:00
The repo root contains the HF-loadable checkpoint files (config.json, tokenizer files, generation_config.json, and *.safetensors). Training-only artifacts such as optimizer state, scheduler state, trainer state, logs, caches, wandb output, and credentials are intentionally excluded.
See checkpoint_metadata.json and SHA256SUMS for provenance and file checksums.
For more details, refer to:
- Project page: https://sharryxr.github.io/ASIL
- GitHub: https://github.com/sharryXR/ASIL