Data cards: the generator and training code are not public (inference-only code); drop script and seed pointers
Browse files
README.md
CHANGED
|
@@ -71,7 +71,7 @@ Each stage is 3–6k steps (10–36 minutes on a laptop).
|
|
| 71 |
|
| 72 |
Every **Khmer target is real text**, except 2,483 chat sentences written by Gemma 4. Only the **romanized input** varies, by source.
|
| 73 |
|
| 74 |
-
Paper: `<paper-url>`.
|
| 75 |
|
| 76 |
## Configs
|
| 77 |
|
|
@@ -83,7 +83,7 @@ The data is organized by kind, so each source of spelling noise can be used, ins
|
|
| 83 |
| `llm_romanized` | `train_gemrom2` (18,273) | Gemma 4 romanizations of 6,382 real sentences, 4 style prompts, ×3 copies with random casing | 2, 3 (×4 in the mix) | Mini, Base |
|
| 84 |
| `abbreviated` | `ft_v2_abbr3` (139,704), `train_gemrom2_abbr3` (16,302) | chat-like abbreviated copies of `chat_mix` and `llm_romanized` pairs; only the input changes (`tov` → `tv`, `bong` → `bg`) | 2, 3 (Gemma part ×2) | Mini, Base |
|
| 85 |
| `mined_tables_chh` | `ft_rules_chh` (289,820) | new pairs spelled with the table mined from the Chhunneng word list (dictionary-like scheme), alone or with mixed / lazy noise habits | 2, 3 | **Mini** |
|
| 86 |
-
| `mined_tables_chh_wd` | `ft_rules_chhwd` (289,647) | made
|
| 87 |
| `casual` | `ft_casual_noakk` (426,946) | 21.6k spoken-register web sentences, spelled by the generator with stretched letters (`tovvv`, `baaan`) | **3 only** | Mini, Base |
|
| 88 |
| `synthetic_heldout` | `train_v2_s3.dev` (12,232) | synthetic held-out pairs monitored during training (validation split) | – | – |
|
| 89 |
|
|
@@ -134,7 +134,7 @@ casual = load_dataset("samdei/samdei-type-finetuning-data", "casual", split="tra
|
|
| 134 |
abbr = load_dataset("samdei/samdei-type-finetuning-data", "abbreviated", split="train")
|
| 135 |
```
|
| 136 |
|
| 137 |
-
To rebuild a mix exactly, read `recipes.json` and concatenate its files in order, with repeats.
|
| 138 |
|
| 139 |
## How the pairs were made
|
| 140 |
|
|
@@ -144,20 +144,20 @@ To rebuild a mix exactly, read `recipes.json` and concatenate its files in order
|
|
| 144 |
3. it spells the piece in a persona (formal 8%, careful 32%, mixed 40%, lazy 20%);
|
| 145 |
4. it adds noise (glued or split words, dropped vowels, swaps, typos), English chat words and place names.
|
| 146 |
- **`llm_romanized`.** Gemma 4 12B, run locally, romanized real sentences "the way Cambodians type on Messenger", 10 sentences per call. The Khmer side stays the original sentence, so the target is always correct, and Gemma's looser spelling acts as human-like noise on the input. 97% of the outputs passed our filters.
|
| 147 |
-
- **`mined_tables_*`.**
|
| 148 |
-
- **`abbreviated`**
|
| 149 |
- It covers 40% of `chat_mix` pairs and all `llm_romanized` pairs.
|
| 150 |
- Operations: consonant skeleton (`bong` → `bg`), dropped vowels, first + last letter (`jong` → `jg`), dropped digraph h (`khlang` → `klang`), dropped cluster member.
|
| 151 |
- Vowels are dropped only from one-syllable words among the 1,000 most frequent.
|
| 152 |
- An abbreviation that the chat lexicon reserves for one word (`nh` for ខ្ញុំ) is never produced for another.
|
| 153 |
-
- **`casual`.** 21.6k HPLT and FineWeb-2 sentences that contain one of 13 spoken-register words (ហ្នឹង, ចឹង, អី, តោះ, …) as a token; misordered PDF text is dropped. They are spelled with stretched letters by the generator (formal 5%, careful 25%, mixed 35%, lazy 15%, Chhunneng table with mixed habits 20%
|
| 154 |
|
| 155 |
**Test separation.**
|
| 156 |
- Every sentence of every evaluation set, including Samdei Type Bench and its chat context, is banned in the generator.
|
| 157 |
- The rule miner never sees a test row or a word-level test target.
|
| 158 |
- No Samdei Type Bench message equals a training target. The only shared strings are common single words.
|
| 159 |
|
| 160 |
-
**Reproducibility.**
|
| 161 |
|
| 162 |
## Limitations
|
| 163 |
|
|
|
|
| 71 |
|
| 72 |
Every **Khmer target is real text**, except 2,483 chat sentences written by Gemma 4. Only the **romanized input** varies, by source.
|
| 73 |
|
| 74 |
+
Paper: `<paper-url>`. Inference code: [github.com/Samdei-Labs/samdei-type](https://github.com/Samdei-Labs/samdei-type). Test set: [`samdei/samdei-type-bench`](https://huggingface.co/datasets/samdei/samdei-type-bench).
|
| 75 |
|
| 76 |
## Configs
|
| 77 |
|
|
|
|
| 83 |
| `llm_romanized` | `train_gemrom2` (18,273) | Gemma 4 romanizations of 6,382 real sentences, 4 style prompts, ×3 copies with random casing | 2, 3 (×4 in the mix) | Mini, Base |
|
| 84 |
| `abbreviated` | `ft_v2_abbr3` (139,704), `train_gemrom2_abbr3` (16,302) | chat-like abbreviated copies of `chat_mix` and `llm_romanized` pairs; only the input changes (`tov` → `tv`, `bong` → `bg`) | 2, 3 (Gemma part ×2) | Mini, Base |
|
| 85 |
| `mined_tables_chh` | `ft_rules_chh` (289,820) | new pairs spelled with the table mined from the Chhunneng word list (dictionary-like scheme), alone or with mixed / lazy noise habits | 2, 3 | **Mini** |
|
| 86 |
+
| `mined_tables_chh_wd` | `ft_rules_chhwd` (289,647) | made the same way from the same sources, but spelled 70% with the Chhunneng table and 30% with the table mined from Wikidata names (official, name-style scheme) | 2, 3 | **Base** |
|
| 87 |
| `casual` | `ft_casual_noakk` (426,946) | 21.6k spoken-register web sentences, spelled by the generator with stretched letters (`tovvv`, `baaan`) | **3 only** | Mini, Base |
|
| 88 |
| `synthetic_heldout` | `train_v2_s3.dev` (12,232) | synthetic held-out pairs monitored during training (validation split) | – | – |
|
| 89 |
|
|
|
|
| 134 |
abbr = load_dataset("samdei/samdei-type-finetuning-data", "abbreviated", split="train")
|
| 135 |
```
|
| 136 |
|
| 137 |
+
To rebuild a mix exactly, read `recipes.json` and concatenate its files in order, with repeats. Training starts from the stage-1 checkpoints, [`samdei/samdei-type-base-pretrained`](https://huggingface.co/samdei/samdei-type-base-pretrained) and [`samdei/samdei-type-mini-pretrained`](https://huggingface.co/samdei/samdei-type-mini-pretrained).
|
| 138 |
|
| 139 |
## How the pairs were made
|
| 140 |
|
|
|
|
| 144 |
3. it spells the piece in a persona (formal 8%, careful 32%, mixed 40%, lazy 20%);
|
| 145 |
4. it adds noise (glued or split words, dropped vowels, swaps, typos), English chat words and place names.
|
| 146 |
- **`llm_romanized`.** Gemma 4 12B, run locally, romanized real sentences "the way Cambodians type on Messenger", 10 sentences per call. The Khmer side stays the original sentence, so the target is always correct, and Gemma's looser spelling acts as human-like noise on the input. 97% of the outputs passed our filters.
|
| 147 |
+
- **`mined_tables_*`.** A rule miner learns per-slot spelling tables (onset, vowel, coda, glide, context) by Viterbi alignment and hard EM from public (Khmer word, romanization) lists: the Chhunneng word list (24.9k pairs, MIT) and 1.6k Wikidata person and place names (CC0). Words that are targets of a word-level evaluation set are removed first. The tables are in [`samdei/samdei-type-resources`](https://huggingface.co/datasets/samdei/samdei-type-resources). The pairs here are new sentences (HPLT, FineWeb-2, ALT, chat sentences) spelled with them.
|
| 148 |
+
- **`abbreviated`.**
|
| 149 |
- It covers 40% of `chat_mix` pairs and all `llm_romanized` pairs.
|
| 150 |
- Operations: consonant skeleton (`bong` → `bg`), dropped vowels, first + last letter (`jong` → `jg`), dropped digraph h (`khlang` → `klang`), dropped cluster member.
|
| 151 |
- Vowels are dropped only from one-syllable words among the 1,000 most frequent.
|
| 152 |
- An abbreviation that the chat lexicon reserves for one word (`nh` for ខ្ញុំ) is never produced for another.
|
| 153 |
+
- **`casual`.** 21.6k HPLT and FineWeb-2 sentences that contain one of 13 spoken-register words (ហ្នឹង, ចឹង, អី, តោះ, …) as a token; misordered PDF text is dropped. They are spelled with stretched letters by the generator (formal 5%, careful 25%, mixed 35%, lazy 15%, Chhunneng table with mixed habits 20%). No Akkhara-derived spelling rules are used.
|
| 154 |
|
| 155 |
**Test separation.**
|
| 156 |
- Every sentence of every evaluation set, including Samdei Type Bench and its chat context, is banned in the generator.
|
| 157 |
- The rule miner never sees a test row or a word-level test target.
|
| 158 |
- No Samdei Type Bench message equals a training target. The only shared strings are common single words.
|
| 159 |
|
| 160 |
+
**Reproducibility.** These files are the exact pairs the released models were trained on, so training can be repeated without regenerating them. The data generator and the training code are not public; the released code ([github.com/Samdei-Labs/samdei-type](https://github.com/Samdei-Labs/samdei-type)) is for inference only.
|
| 161 |
|
| 162 |
## Limitations
|
| 163 |
|