Tags: RLWRLD/RLDX-1
Tags
Release v1.0.2: DROID deploy configuration placeholders The parameters.py example in the DROID deploy guide now uses <...> placeholders for every setup-specific value (IPs, sudo password, robot serial number, camera IDs, Ubuntu Pro token), with a note to keep the filled-in file local.
RLDX-1 v1.0.1 Fine-tuning default alignment + single-source versioning. - state_dropout_prob: 0.0 -> 0.3 - tune_top_llm_layers -> 4 (model config aligned with CLI) - color_jitter on by default - max_steps: 10000 -> 30000 - benchmark/midtrain scripts pin prior behavior with explicit flags - versioning: single source of truth (rldx.__version__), bumped to 1.0.1
fix(launcher): resolve backbone from config.json to prevent silent re… …-init from RLDX-1-PT (#22) Fine-tuning from RLWRLD/RLDX-1-PT without --backbone-path silently rebuilt the backbone against vanilla Qwen/Qwen3-VL-8B-Instruct (vocab 151 936), discarded the checkpoint's embed_tokens / lm_head (vocab 153 720) and re-initialised them randomly — ~1.26 B params (~18 % of the 6.9 B model), with no visible warning. Cause: RLDX-1-PT ships HF-style config.json (with model_name set) but no experiment_cfg/config.yaml; resolve_backbone_path only consulted the yaml and fell through to the vanilla-Qwen default. resolve_backbone_path now resolves in this order: 1. --backbone-path CLI override 2. experiment_cfg/config.yaml snapshot's model_name 3. <base_model_path>/config.json's model_name (logged) 4. Qwen/Qwen3-VL-8B-Instruct ONLY when base_model_path is None (genuine from-scratch). Checkpoints whose backbone cannot be identified now raise instead of silently re-initialising against the wrong vocabulary. assembly._apply_cli_model_overrides / assemble_run_config threaded base_model_path through; dead ckpt_config={} stub dropped. Verified end-to-end with a real RLDX-1-PT load: before "2 shape-mismatched, 3 missing" -> after "0 / 0 / 0". Co-authored-by: jaehyun <jaehyun@ip-172-31-40-242.ec2.internal>