| base_model: Qwen/Qwen3-4B-Instruct-2507 | |
| library_name: peft | |
| tags: | |
| - peft | |
| - lora | |
| - prime-rl | |
| - rlm-rlvr | |
| - qwen3 | |
| # Qwen3-4B RLM RLVR Depth-1 LoRA Adapter | |
| LoRA adapter from the first rank/LR ablation run. | |
| - Base model: `Qwen/Qwen3-4B-Instruct-2507` | |
| - Run id: `rlm-rlvr-qwen3-4b-depth1-llmonly-r4-a8-lr5e-7-s150` | |
| - Prompt variant: `sanjaya_text_depth1_llm_only_v1` | |
| - Runtime depth: depth-1 LLM-only orchestration (`max_depth = 0`, plain `llm_query` subcalls enabled) | |
| - LoRA rank: 4 | |
| - LoRA alpha: 8 | |
| - Learning rate: 5e-7 | |
| - Training steps: 150 | |
| - Final adapter source: `run_default/broadcasts/step_150` | |
| The `run_configs/` directory contains the exact trainer and orchestrator TOML files saved with the run. | |