# MyML > Run real GGUF language models locally on your GPU inside the browser with WebGPU—and open the hood to steer thoughts live, inspect every transformer layer, edit weights, and train LoRA adapters with zero servers and zero API keys. - **Canonical URL:** https://myml.deeflect.com/ - **Source Code (MIT):** https://github.com/deeflect/myml - **Category:** Browser-based WebGPU LLM engine, mechanistic interpretability workbench, and local ML sandbox - **Privacy:** 100% local execution. Models are downloaded from Hugging Face (or loaded from local `.gguf` files) directly into the browser's Origin Private File System (OPFS) and run on the user's GPU via WebGPU. No prompts, replies, or weights ever leave the device. - **Pricing:** Free, no login or API key required. ## Supported Models & Hugging Face GGUF Loader MyML includes a pure TypeScript + core-WGSL WebGPU GGUF inference engine supporting Llama, Qwen2, Qwen3 (QK-norm), and SmolLM3 (NoPE every 4th layer) architectures across `Q4_K_M`, `Q5_K_M`, `Q6_K`, `Q8_0`, `Q4_0`, `F16`, and `BF16` quantizations: - **Qwen3 0.6B** (`Q4_K_M`, ~378 MB, fastest starter preset, supports `` trace toggle) - **Llama 3.2 1B Instruct** (`Q4_K_M`, ~770 MB) - **Qwen3 1.7B** (`Q4_K_M`, ~1.03 GB, supports `` trace toggle) - **SmolLM3 3B** (`Q4_K_M`, ~1.78 GB, NoPE + `` trace toggle) - **Any compatible Hugging Face GGUF:** Paste an `owner/repo`, Hugging Face URL, direct `.gguf` link, or search Hugging Face directly with automatic pre-download architecture and quantization compatibility checks, or open a local `.gguf` file from disk. ## Core Workspaces ### 1. Chat & Live Dials - **Behavior Dials (CAA Activation Steering):** Drag continuous dials (`Directness`, `Skepticism`, `Corporate caution`, `Lateral thinking`, or custom dials taught from contrastive prompt pairs) across specific transformer layer bands with optional L2 residual-norm preservation. - **Token Multiverse (Click-to-Fork):** Click any word in a reply to inspect the top 5 runner-up tokens and probabilities, then rewind and stream an alternate continuation from that exact token. - **Dial Spectrum Sweep:** Stream the latest prompt side-by-side at `−1.5`, `0.0`, and `+1.5` of the active dial in 3 columns. - **Live Mood Stethoscope (Passive Meters):** Project the live middle-layer residual stream onto every dial in real time—even dials set to `0.0`. - **Conditional Dials (CAST):** Trigger one dial automatically only when another concept direction fires above a chosen threshold. - **Think Mode (``):** Toggle hybrid-thinking scratchpads on supported models (`Qwen3`, `SmolLM3`) with clickable tokens inside the collapsible thought trace. - **Bake into Model (Weight Abliteration):** Permanently orthogonalize or bake active steering directions into `attn_output` and `ffn_down` GPU weights, or export as a `llama.cpp`-compatible Control Vector `.gguf` or shareable `#patch=` URL. ### 2. Inside (X-Ray & Live Surgery) - **Logit Lens:** See the model's evolving next-word prediction and certainty across every transformer layer for any clicked token. - **Direct Logit Attribution (DLA):** Measure whether each layer's Attention or MLP pushed the chosen word up or down. - **Live Architecture Surgery:** Skip any layer, loop a layer twice, or mute individual attention heads live without reloading weights. ### 3. Lab (10 Mechanistic Interpretability & Editing Tools) - **Find:** Concept Neurons (MLP down-projection search with Mute / Normal / Boost), Weight Axes (SVD extraction), and Word Math (`token_embd` vector arithmetic). - **Edit:** Activation Transplant (A→B causal patching across layer bands), Compress (live 8/4/3/2/1-bit weight crushing + Wanda pruning), Blend (2D XY Personality Pad with bilinear 4-corner dial mixing, SLERP/Linear blending, and TIES/DARE LoRA merging), and Features (Single-Layer Sparse Autoencoder clamping). - **Generate:** Sampling workbench (Entropix adaptive sampling, Min-P, Top-nσ, DRY repetition penalty, XTC, plus Word Ban `−∞` and Word Magnet `+4` logit bias) and Settings Physics (`RoPE time-warp` positional scaling and `Thought fever` residual activation noise). - **Train:** Dataset Surprise Heatmap & Perplexity Eval (`.jsonl` / `.txt`), plus live in-browser Rank-4 LoRA fine-tuning with AdamW, GPU weight fusion, and `.safetensors` export. ### 4. Learn (69 ELI5 Topics) Every slider, toggle, metric, and ML concept in MyML includes an inline `?` ELI5 tooltip linked to a 69-topic searchable Learn guide explaining each mechanism in plain English with everyday analogies. ## Persistence & Export - Automatic OPFS workspace auto-save and named snapshots per model. - Portable `.myml.json` workspace backup/restore, `llama.cpp` Control Vector `.gguf` export/import, LoRA `.safetensors` export, and compact `#patch=` share links.