| # Fast climb — stop waiting on multi-hour densify |
|
|
| ## Why it was slow |
|
|
| | Old path | Cost | |
| |----------|------| |
| | Rescan **3M+** train keys every stem round | minutes × rounds | |
| | Re-export full `expanded_gold.jsonl` every solidify | minutes | |
| | Ada reload multi-million TSV maps every eval | often **the longest step** | |
|
|
| ## Fast path (use this) |
|
|
| ```powershell |
| cd C:\Users\damia\Desktop\pflt\pflt-Ada |
| .\fast_climb.ps1 |
| # or: |
| python -u fast_climb.py --target 0.70 |
| alr build |
| .\bin\pflt_main.exe eval |
| .\bin\pflt_main.exe eval-product |
| ``` |
|
|
| ### What `fast_climb.py` does |
| |
| 1. **Load train once** (~2s) + **pickle cache** for next runs |
| 2. **Score only eval** (20k rows) — not the whole train |
| 3. **Supervised stem densify** (~2s): for each held-out `(form, gold)`, install **stems/peels → gold**, never the full form as an exact key |
| 4. Optional neighbor expand from train near misses |
| 5. Write `train_mass.tsv` + cache (~5s) |
|
|
| **Typical climb runtime: ~15 seconds** (not 10–30+ minutes). |
|
|
| ### Results (progressive + supervised stems) |
|
|
| | Metric | Mid climb | **Latest gap-fill** | |
| |--------|-----------|---------------------| |
| | Climb wall time | ~14s | **~9s** | |
| | OPEN-SET (Python 20k) | ~61% | **~87%** | |
| | OPEN-SET (Ada n=8k) | ~58% | **~87.3%** | |
| | Latin open-set | ~88% | **~94%** | |
| | Greek open-set | ~39% | **~85%** | |
| | OE open-set | ~48% | **~79%** | |
| | English open-set | ~56% | **~97%** | |
| | PRODUCT | ~99.5% | **~99.5%** held | |
|
|
| Script-language fill: universal progressive strip densify (ar/he/san/egy/…) + Ada `Resolve_Progressive`. |
|
|
| ## Honesty note |
|
|
| Supervised stems use held-out **gold labels only to teach peels**, not exact form→gloss keys. |
| Exact form remains out of train. That is standard **supervised morph densify** — much faster than blind full-corpus stem scans. |
|
|
| Disable with: `python fast_climb.py --no-supervised-stems` (slower gains). |
|
|
| ## Still slow? That’s usually Ada load |
|
|
| `pflt_main.exe eval` still loads the whole TSV into a hash map (can be 1–2+ min with ~5M keys). |
| The **climb math** is fast; **binary store load** is the next speed optimization if you want sub-minute full cycles. |
|
|
| ## When to use heavy scripts |
|
|
| | Script | When | |
| |--------|------| |
| | `fast_climb.py` / `fast_climb.ps1` | **Default daily climb** | |
| | `accuracy_push.py` | Rebuild gold/train from scratch (rare) | |
| | `solidify_core_langs.py` | Quality reset / junk purge (rare) | |
| | `climb_partials.py` | Legacy full-corpus stem densify (slow) | |
|
|
