remove card
Browse files
README.md
DELETED
|
@@ -1,35 +0,0 @@
|
|
| 1 |
-
---
|
| 2 |
-
library_name: transformers
|
| 3 |
-
pipeline_tag: text-generation
|
| 4 |
-
tags:
|
| 5 |
-
- african-languages
|
| 6 |
-
- code-switching
|
| 7 |
-
- gemma3
|
| 8 |
-
- sft
|
| 9 |
-
---
|
| 10 |
-
|
| 11 |
-
> **Private development checkpoint (baseline "sft1"), not for release.** Read "Data caveats" before any use of this model or any claim made with it.
|
| 12 |
-
|
| 13 |
-
# csw sft1: code-switching and idiomatic rewriting for 37 African languages (270M)
|
| 14 |
-
|
| 15 |
-
Start point: `BeardedMonster/gemma-3-270m-translate-rl2c` (English to African translation, GRPO phase 2c). Its card states: no NLLB signal, no TranslatePsy data, no AfriScience-MT data in its lineage; its older SFT data include non-commercial and unstated-licence sources.
|
| 16 |
-
Full fine-tuning (fp32 weights, bf16 autocast, AdamW lr 5e-5, 2 epochs, 262,144 tokens per optimiser step, 2,305 steps). Best validation loss **1.2508** (step 2,305; 1.5223 at step 250).
|
| 17 |
-
|
| 18 |
-
## Tasks and format
|
| 19 |
-
Prompts (ISO tags, base language first): `<hm> <yor>+<eng>: {text}` (English or native text to code-switched), `<ood> <yor>: {text}` (to informal native text with idioms), `<ood> <yor>+<eng>: {hm text}`.
|
| 20 |
-
Completion: `<yor>[<switch>]<dev>device one | device two</dev> text`. Every tag is one vocabulary token (renamed `<unusedN>` slots; vocabulary size unchanged at 262,144).
|
| 21 |
-
Replay of the original task: `Translate the text to Yoruba: {english}` -> `yor {text}`.
|
| 22 |
-
|
| 23 |
-
## Training data (829,538 train examples, 299.8 M tokens)
|
| 24 |
-
* eng_ood 189,707; m_ood 189,271; eng_hm 150,144; m_hm 148,779; hm_ood 138,357: text generated by `google/gemma-4-31b-it` (FP8) on four task cards, then filtered by rules (English leak, switch ratio, device count), by two LLM judges (Gemini 3.8 Flash, GPT-5.6 luna; Batch API) on about 22% of rows, and with the judges' corrected copies added next to the originals (a field a trusted judge called major/critical is withheld from the original). Rows not judged were not checked by a judge.
|
| 25 |
-
* eng_m replay 13,280 examples (13,533 incl. validation) of human-translated English to African pairs.
|
| 26 |
-
|
| 27 |
-
## Data caveats (important)
|
| 28 |
-
* **The replay set was NOT decontaminated against benchmark dev/devtest/test text.** It includes 3,600 pairs from `dsfsi/afriscience_mt` **dev** references (against the owner's rule of 2026-10-28), and about 9,000 pairs taken from `BeardedMonster/sft-all`, `opus-100` and `African-Languages-Lab/multi-open`, which were not run through the strict filter (the 2026-10-26 audit found 95% of FLORES+ devtest English in the older SFT text). It is 1.6% of the training examples. Do not evaluate this model's translation on AfriScience-MT or on benchmarks that these sources may contain.
|
| 29 |
-
* Replay licences: multi-open "other"; opus-100 mixed; Tatoeba CC-BY-2.0-FR; Gamayun CC-BY-4.0; AfriScience Apache-2.0. Not cleared for commercial use.
|
| 30 |
-
* Not used: NLLB, TranslatePsy, FineTranslations, mined or non-commercial files.
|
| 31 |
-
* The generated code-switch text is machine output (a 31B model) and 78% of it was never checked by a judge; languages with little judge coverage have little evidence of quality.
|
| 32 |
-
* The curated test split (4,346 examples) was drawn from judged rows with no major or critical error; test-split rows share no English source sentence with the training rows.
|
| 33 |
-
|
| 34 |
-
## Status
|
| 35 |
-
Evaluation (chrF on the curated test split, English-leak rate, device-in-text rate) is pending; this card will be updated with the numbers.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|