ryanontheinside/experimental-acestep1.5-v1

Experimental

Experimental-style generation. Covers: experimental electronic, glitchy textures, abstract rhythms, evolving sound design; experimental noise, harsh textures, manipulated samples, dense layering; and experimental ambient, processed field recordings, drones, sparse rhythms.

Quick start

• Base model: AceStep v1.5 Turbo (2B)
• Trigger word: roti-3xpr1mtl (must appear in the prompt to activate)
• Recommended strength: 1.0
• Tested strength range: 0.3 – 1.0
• Recommended steps: 8
• Recommended guidance: 7.0
• Recommended shift: 3.0
• Sample rate: 48000 Hz

Example prompts

• experimental electronic, glitchy textures, abstract rhythms, evolving sound design (110 BPM, C minor)
• experimental noise, harsh textures, manipulated samples, dense layering (95 BPM, D minor)
• experimental ambient, processed field recordings, drones, sparse rhythms (60 BPM, G minor)

Training

• Rank 64, alpha 128, dropout 0.1
• Target modules: q_proj, k_proj, v_proj, o_proj
• 500 epochs / 7500 steps, batch 1, optimizer adamw_fused, lr 0.0003, warmup 100
• Dataset: 15 synthetic tracks, 3485s total, instrumental fraction 0.2

Files

• experimental-v1.safetensors — PEFT LoRA weights (88,130,248 bytes, sha256 83590b8c9f7921f26225d52469ae99d29e5d5e8a6abe627d868113e6f0938262)
• experimental-v1.metadata.json — machine-readable sidecar (schema v1)
• adapter_config.json — PEFT adapter config

Full metadata: see experimental-v1.metadata.json.