Add README
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Showing 1 changed file with +128 and −0
added
README.md
+128 −0
@@ -0,0 +1,128 @@ | ||
| 1 | +<!-- Author: Simon-Pierre Boucher — contact@spboucher.ai --> | |
| 2 | + | |
| 3 | +<div align="center"> | |
| 4 | + | |
| 5 | +# 🔥 Forge Studio | |
| 6 | + | |
| 7 | +### The Instruments of LLM training — a native macOS cockpit for [Forge](https://github.com/spboucher-ai/forge) | |
| 8 | + | |
| 9 | +**Train language models from scratch on Apple Silicon without ever opening a terminal.** | |
| 10 | +Pick a dataset, shape an architecture, hit Start — and watch the loss fall on a | |
| 11 | +silky live dashboard. | |
| 12 | + | |
| 13 | +Swift + SwiftUI + Swift Charts · macOS 14+ · Apple Silicon · zero third-party dependencies | |
| 14 | + | |
| 15 | +</div> | |
| 16 | + | |
| 17 | +--- | |
| 18 | + | |
| 19 | +## ✨ What it is | |
| 20 | + | |
| 21 | +[Forge](https://github.com/spboucher-ai/forge) is a from-scratch C++20 + Metal LLM | |
| 22 | +training framework: its own tensors, autograd, flash attention, Muon optimizer, | |
| 23 | +MoE, QAT, and the `.forge` zero-copy weight format. **Forge Studio is its native | |
| 24 | +GUI companion** — the whole train-a-model workflow, end to end, as a first-party-feeling | |
| 25 | +Mac app: | |
| 26 | + | |
| 27 | +1. **Prepare data** — TinyStories or streamed Hugging Face mixtures | |
| 28 | + (FineWeb-Edu, DCLM, Cosmopedia, research presets), with a live console | |
| 29 | +2. **Design a model** — every Forge config field, from `n_layers` to | |
| 30 | + DeepSeek-style MoE routing, with live validation | |
| 31 | +3. **Train** — launch, monitor, stop; runs survive app crashes without lying | |
| 32 | +4. **Understand** — a loss dashboard built to the TensorBoard/W&B standard, native | |
| 33 | +5. **Compare** — multi-run overlays on the *tokens* axis, the honest one | |
| 34 | +6. **Use the model** — generate and evaluate from any checkpoint, in-app | |
| 35 | + | |
| 36 | +Forge Studio never reimplements training. It drives the real `forge` binary and | |
| 37 | +reads its structured metrics — what you see is exactly what the framework did. | |
| 38 | + | |
| 39 | +## 📊 The dashboard | |
| 40 | + | |
| 41 | +The centerpiece. Built on Swift Charts, fed by an actor-isolated metrics store: | |
| 42 | + | |
| 43 | +- **Raw + smoothed loss** — low-opacity raw train line under a bias-corrected EMA | |
| 44 | + (TensorBoard semantics, 0→0.99 slider), val loss as connected points | |
| 45 | +- **Hover crosshair** with a full callout: step, train, EMA, perplexity, val | |
| 46 | +- **Pinch-zoom & pan** on X, double-click to reset, and a **"follow live ⏵" pill** | |
| 47 | + that re-pins the window to incoming data after you've panned away mid-run | |
| 48 | +- **Best-val marker** — dashed rule + `best val 2.146 @ step 999` annotation | |
| 49 | +- **Log/linear Y**, secondary charts for LR schedule (watch warmup + cosine/WSD | |
| 50 | + actually happen), tokens/sec, grad norm with the `grad_clip` threshold drawn | |
| 51 | +- **Scale discipline**: raw data is never discarded; the UI reads LTTB-downsampled | |
| 52 | + snapshots (~2× pixel width). Measured: 200,000 CSV rows ingest in ~1.1 s and | |
| 53 | + snapshot to chart width in <250 ms — hover stays hitch-free on 100k-step runs | |
| 54 | + | |
| 55 | +## 🛡️ Runs that can't lie | |
| 56 | + | |
| 57 | +- Single-writer **state machine** (`queued → launching → running → … → finished | | |
| 58 | + failed | stopped`) with an explicit legal-transition table — illegal transitions | |
| 59 | + are unrepresentable | |
| 60 | +- **Atomic persistence** (temp-file-then-rename) for the registry: `kill -9` the | |
| 61 | + app whenever you like | |
| 62 | +- **Crash recovery**: runs persisted as active at launch are truthfully resolved — | |
| 63 | + including the case where the forge process is *still alive* — with an honest | |
| 64 | + explanation and the recovery path (`ckpt_latest.bin`) | |
| 65 | +- **Watchdog**: a "possibly stalled" badge when metrics stop flowing for 30 s | |
| 66 | +- Stop is SIGTERM and the UI tells you exactly what that means (Forge doesn't | |
| 67 | + checkpoint on signals; you lose at most `checkpoint_every` steps) | |
| 68 | +- Local **notifications** with the final loss when a run finishes or fails | |
| 69 | + | |
| 70 | +## 🧪 Ground truth, tested | |
| 71 | + | |
| 72 | +`RESEARCH.md` documents the full Forge contract (config schema, CLI, `log.csv` | |
| 73 | +grammar, checkpoint/resume, signal behavior) extracted from the source. The test | |
| 74 | +suite enforces it: | |
| 75 | + | |
| 76 | +- `ForgeConfig` round-trips the **real** `configs/*.json` from the forge repo, | |
| 77 | + byte-compatible field names | |
| 78 | +- The Swift param-count formula matches `forge info` for every shipped config | |
| 79 | +- CSV/stdout parsers are exercised against real and mutated lines (legacy | |
| 80 | + 6-column headers, garbage, truncation) | |
| 81 | +- LTTB invariants (endpoints exact, monotonic X), EMA bias correction, | |
| 82 | + state-machine table, LR-preview math vs `scheduler.h`, 200k-point stress test | |
| 83 | + | |
| 84 | +## 🚀 Quick start | |
| 85 | + | |
| 86 | +```bash | |
| 87 | +# 1. Build Forge (once) | |
| 88 | +git clone https://github.com/spboucher-ai/forge && cd forge | |
| 89 | +cmake -B build && cmake --build build -j | |
| 90 | + | |
| 91 | +# 2. Run Forge Studio | |
| 92 | +# — from the DMG on the Releases page, or from source: | |
| 93 | +git clone https://github.com/spboucher-ai/forge-studio && cd forge-studio | |
| 94 | +./scripts/package-app.sh release && open dist/ForgeStudio.app | |
| 95 | +``` | |
| 96 | + | |
| 97 | +In **Settings**: point to `forge/build/forge` (validated live via `forge info`) and | |
| 98 | +pick a workspace. Then **Datasets → New dataset** (TinyStories at vocab 4096 takes | |
| 99 | +a few minutes), **＋ New run**, pick the `gpt-10m` preset, Start — first loss point | |
| 100 | +lands within seconds. | |
| 101 | + | |
| 102 | +## 🏗️ Architecture | |
| 103 | + | |
| 104 | +| Layer | What lives there | | |
| 105 | +|:--|:--| | |
| 106 | +| `Models/` | `ForgeConfig` (Codable mirror of every Forge field + validation + derived math), `Run` (state machine), `MetricPoint`, `Dataset` (bin-header readers) | | |
| 107 | +| `Services/` | `ProcessRunner` (actor, incremental line streaming), `LogParser` (header-driven CSV + stdout events), `MetricsStore` (actor, incremental tail, LTTB snapshots), `RunStore` (atomic registry), `RunSupervisor`, `ForgeBinaryLocator`, `SystemInfo` | | |
| 108 | +| `Charts/` | `TrainingChartView` (the interactive dashboard), `Downsampler` (LTTB), `Smoothing` (bias-corrected EMA) | | |
| 109 | +| `Views/` | Navigation shell, run dashboard, New Run editor (presets, derived panel, LR preview), Compare, Datasets, Checkpoints/Generate/Eval, Settings | | |
| 110 | +| `scripts/` | `package-app.sh` (SwiftPM → .app), `generate-icon.sh`, `notarize.sh` (Developer ID + hardened runtime + DMG + notarytool + stapler) | | |
| 111 | + | |
| 112 | +Swift Concurrency throughout: the process I/O and metrics paths never touch the | |
| 113 | +main thread; the UI observes debounced snapshots. No third-party packages. | |
| 114 | + | |
| 115 | +## 🗺️ Roadmap | |
| 116 | + | |
| 117 | +- Linked axes across secondary charts + synchronized crosshair | |
| 118 | +- Clickable checkpoint annotations on the loss chart | |
| 119 | +- Exhaustive form for every architecture-variant knob (today: JSON import covers them) | |
| 120 | +- PNG/CSV chart export, run queue for sequential trainings | |
| 121 | +- `.forge` repo browser (weight history as commits — the format already does git-style deltas) | |
| 122 | + | |
| 123 | +## 👤 Author | |
| 124 | + | |
| 125 | +**Simon-Pierre Boucher** — contact@spboucher.ai | |
| 126 | + | |
| 127 | +Built alongside [Forge](https://github.com/spboucher-ai/forge). If Forge is the | |
| 128 | +engine, Studio is the cockpit. | |
| 129 | ||