| --- |
| private: true |
| tags: [polyplex, personaplex, korean, full-duplex] |
| --- |
| |
| # PolyPlex Korean, Phase 1.0 and 1.5 adapters |
|
|
| Backup of the two Phase 1 stages of PolyPlex, which adds Korean to a frozen PersonaPlex 7B from text only. |
| These are adapter slices, not full models: each holds the appended Korean text rows and the session expert |
| (a correction on the FFNs of layers 8 to 31, rank 256). They load onto the vocabulary-expanded base |
| checkpoint (`expanded_v2`, 32,000 + 10,448 tokens) with the PolyPlex code |
| (`polyplex_research.utils.assemble.Phase1Assemble.from_run`). |
|
|
| | Folder | Stage | Trained on | Code | |
| |---|---|---|---| |
| | `phase1.0/` | Dense text (no PAD frames), 26,000 steps | About 510M tokens of Korean web text (HAERAE KOREAN-WEBTEXT) and Korean Wikipedia | branch `experiment/two-stage-dense-duplex` (commit 2e00113) | |
| | `phase1.5/` | Duplex dialogue layout with codec silence, 6,000 steps, started from `phase1.0` | 384,000 of 602,677 dialogue samples (AI Hub multisession dialogue 011 and 141, colloquial translation sentences) | same branch | |
|
|
| Each folder holds the final adapter (`config.json`, `text_rows.safetensors`) and the last trainer checkpoint |
| (`checkpoint-26000/`, `checkpoint-6000/`) with optimizer, scheduler and RNG state for resuming. |
|
|
| ## Results of `phase1.5/` (silent condition unless stated) |
|
|
| | Measure | Value | |
| |---|---| |
| | Korean word loss, 011 validation (8,348 utterances, training sentences removed) | 4.136 (PPL 62.6) | |
| | Korean word loss, 141 dialogue validation (2,000 dialogues) | 3.738 | |
| | Same-meaning comparison, 200 pairs: Korean / English (original model) | 4.067 / 3.835 | |
| | Timing cost on 011 (PAD/EPAD loss over the original model) | +0.056 | |
| | Korean word loss with real speech (KsponSpeech validation) | 9.322 | |
|
|
| Known limits: Korean does not carry over to the condition where the assistant channel holds real speech, and |
| with the session expert on the depth decoder's code prediction is broken (assistant code loss 7.87 against |
| 4.38 with the expert off), so this adapter does not speak Korean. The full records are the ResearchTree pull |
| requests #17 (this stage) and #18, #19, #20 (later attempts) of `latentforge/PolyPlex`. |
|
|
| The training data licenses have not been reviewed for redistribution, which is why this repository is private. |
|
|
