AntimLabs/Pi-Lumine-4B-FlowMatching

Pi-Lumine 4B — Flow-Matching Action Decoder for Elden Ring

A Pi0.5-style flow-matching action decoder trained on top of a frozen Qwen3.5-4B VLM backbone.

Architecture

• Base VLM: Qwen/Qwen3.5-4B (frozen, not included — downloaded at runtime)
• Action Decoder: FiLM-conditioned transformer with cross-attention to VLM hidden states
  • 2 decoder layers, VLM dim 2560 → decoder dim 1024, 8 attention heads
  • Projection layers decouple decoder from VLM hidden size
  • Instruction-conditioned via AdaptiveRMSNorm (FiLM)
  • Sinusoidal time embedding for flow matching
  • ~64M trainable parameters
• Action Space: 6 steps x 20 dims (4 sticks + 16 buttons per step)
• Training: Flow matching with Euler ODE integration at inference

Files

• action_decoder.pt — Trained action decoder weights
• decoder_config.json — Architecture and tokenizer config
• tokenizer.json / tokenizer_config.json — Tokenizer with special tokens
• chat_template.jinja — Chat template
• processor_config.json — Processor config