RL-unseen-language
Collection
Using RL to elicit context leverage ability of LLMs to learn unseen languages! • 2 items • Updated
original stringlengths 1 658 | eng_translation stringclasses 258
values | data_source stringclasses 2
values | prompt listlengths 1 1 | ability stringclasses 1
value | reward_model dict | extra_info dict | direction stringclasses 26
values | translation stringlengths 1 658 ⌀ | legacy_variety stringclasses 6
values |
|---|---|---|---|---|---|---|---|---|---|
ˈnè=ko ˈjéna ˈá ˈwé kaˈlé ˈnè=ko ˈá riˈká ˈwé tiˈbú-ma oˈlá ˈá oˈtʃ͡êri-ma aˈlé | I do love him very much, I will always take care of him, he will grow. | back-to-school | [
{
"content": "Choguita Rarámuri, a Uto-Aztecan language spoken in the Sierra Tarahumara, a mountainous range in the northern Mexican state of Chihuahua belonging to the Sierra Madre Occidental.\nYour task is to translate from Choguita to English: ˈnè=ko ˈjéna ˈá ˈwé kaˈlé ˈnè=ko ˈá riˈká ˈwé tiˈbú-ma oˈlá ˈá oˈ... | translation | {
"ground_truth": "I do love him very much, I will always take care of him, he will grow.",
"style": "rule"
} | {
"answer": "I do love him very much, I will always take care of him, he will grow.",
"grammar_metadata": [
{
"metadata": {
"Header 1": "",
"Header 2": "",
"Header 3": "",
"Header 4": "",
"source": "/home/zsnajdr/rl-for-low-resource-languages/verl_preprocessing/reso... | Choguita_English-5-sentences | null | null |
amiˈná beˈlá=timi ˈmá oˈtʃéra-ame ˈní-bo aˈlé | Later on when you all will be adults. | back-to-school | [
{
"content": "Choguita Rarámuri, a Uto-Aztecan language spoken in the Sierra Tarahumara, a mountainous range in the northern Mexican state of Chihuahua belonging to the Sierra Madre Occidental.\nYour task is to translate from Choguita to English: amiˈná beˈlá=timi ˈmá oˈtʃéra-ame ˈní-bo aˈlé\n\nTo help with th... | translation | {
"ground_truth": "Later on when you all will be adults.",
"style": "rule"
} | {
"answer": "Later on when you all will be adults.",
"grammar_metadata": [
{
"metadata": {
"Header 1": "",
"Header 2": "",
"Header 3": "",
"Header 4": "",
"source": "/home/zsnajdr/rl-for-low-resource-languages/verl_preprocessing/resources//choguita_raramuri/grammar_... | Choguita_English-5-sentences | null | null |
ka=ni maˈtʃ͡í-ki [ ˈnápi ˈtò-ru ba] | I didn’t know he had been taken | back-to-school | [
{
"content": "Choguita Rarámuri, a Uto-Aztecan language spoken in the Sierra Tarahumara, a mountainous range in the northern Mexican state of Chihuahua belonging to the Sierra Madre Occidental.\nYour task is to translate from Choguita to English: ka=ni maˈtʃ͡í-ki [ ˈnápi ˈtò-ru ba]\n\nTo help with the translati... | translation | {
"ground_truth": "I didn’t know he had been taken",
"style": "rule"
} | {
"answer": "I didn’t know he had been taken",
"grammar_metadata": [
{
"metadata": {
"Header 1": "",
"Header 2": "",
"Header 3": "",
"Header 4": "",
"source": "/home/zsnajdr/rl-for-low-resource-languages/verl_preprocessing/resources//choguita_raramuri/grammar_plain.... | Choguita_English-5-sentences | null | null |
aʔˈlì=ko ˈnè ˈmá biˈlá ˈnápa ˈnì oˈʰkóriri ˈnè wiˈlí ˈnà iˈt͡ʃîpika ba neˈká ba | So then I stood behind/between the trees, hiding, watching. | back-to-school | [
{
"content": "Choguita Rarámuri, a Uto-Aztecan language spoken in the Sierra Tarahumara, a mountainous range in the northern Mexican state of Chihuahua belonging to the Sierra Madre Occidental.\nYour task is to translate from Choguita to English: aʔˈlì=ko ˈnè ˈmá biˈlá ˈnápa ˈnì oˈʰkóriri ˈnè wiˈlí ˈnà iˈt͡ʃîpi... | translation | {
"ground_truth": "So then I stood behind/between the trees, hiding, watching.",
"style": "rule"
} | {
"answer": "So then I stood behind/between the trees, hiding, watching.",
"grammar_metadata": [
{
"metadata": {
"Header 1": "",
"Header 2": "",
"Header 3": "",
"Header 4": "",
"source": "/home/zsnajdr/rl-for-low-resource-languages/verl_preprocessing/resources//chog... | Choguita_English-5-sentences | null | null |
ˈmá ku siˈmí-li maˈnûel, we ra ˈke-o? | Manuel already left, or not? | back-to-school | [
{
"content": "Choguita Rarámuri, a Uto-Aztecan language spoken in the Sierra Tarahumara, a mountainous range in the northern Mexican state of Chihuahua belonging to the Sierra Madre Occidental.\nYour task is to translate from Choguita to English: ˈmá ku siˈmí-li maˈnûel, we ra ˈke-o?\n\nTo help with the transl... | translation | {
"ground_truth": "Manuel already left, or not?",
"style": "rule"
} | {
"answer": "Manuel already left, or not?",
"grammar_metadata": [
{
"metadata": {
"Header 1": "",
"Header 2": "",
"Header 3": "",
"Header 4": "",
"source": "/home/zsnajdr/rl-for-low-resource-languages/verl_preprocessing/resources//choguita_raramuri/grammar_plain.txt... | Choguita_English-5-sentences | null | null |
ˈwé aˈnátʃ͡a-o ka biˈlé iˈsî- nale ba | She endures a lot, she doesn't want to pee. | back-to-school | [
{
"content": "Choguita Rarámuri, a Uto-Aztecan language spoken in the Sierra Tarahumara, a mountainous range in the northern Mexican state of Chihuahua belonging to the Sierra Madre Occidental.\nYour task is to translate from Choguita to English: ˈwé aˈnátʃ͡a-o ka biˈlé iˈsî- nale ba\n\nTo help with the transla... | translation | {
"ground_truth": "She endures a lot, she doesn't want to pee.",
"style": "rule"
} | {
"answer": "She endures a lot, she doesn't want to pee.",
"grammar_metadata": [
{
"metadata": {
"Header 1": "",
"Header 2": "",
"Header 3": "",
"Header 4": "",
"source": "/home/zsnajdr/rl-for-low-resource-languages/verl_preprocessing/resources//choguita_raramuri/gr... | Choguita_English-5-sentences | null | null |
aʔˈlì a ˈjá saˈjè-li [naˈwà-a-tʃ͡i ˈétʃ͡i koriˈmá paˈtʃ͡á bitiˈtʃ͡í baˈkí-li] | And then they felt when the fire bird arrived and when he went inside the house. | back-to-school | [
{
"content": "Choguita Rarámuri, a Uto-Aztecan language spoken in the Sierra Tarahumara, a mountainous range in the northern Mexican state of Chihuahua belonging to the Sierra Madre Occidental.\nYour task is to translate from Choguita to English: aʔˈlì a ˈjá saˈjè-li [naˈwà-a-tʃ͡i ˈétʃ͡i koriˈmá paˈtʃ͡á bitiˈt... | translation | {
"ground_truth": "And then they felt when the fire bird arrived and when he went inside the house.",
"style": "rule"
} | {
"answer": "And then they felt when the fire bird arrived and when he went inside the house.",
"grammar_metadata": [
{
"metadata": {
"Header 1": "",
"Header 2": "",
"Header 3": "",
"Header 4": "",
"source": "/home/zsnajdr/rl-for-low-resource-languages/verl_preproce... | Choguita_English-5-sentences | null | null |
wiˈʰkâ napaˈwí-ka raˈʔìtʃ͡i-li ˈmí reʔˈlé-ki esˈk extsuperscriptêla-tʃ͡i=ko, ˈnè ke siˈné iˈjêna | Many (people) got together to speak over there at the school, but I've never gone. | back-to-school | [
{
"content": "Choguita Rarámuri, a Uto-Aztecan language spoken in the Sierra Tarahumara, a mountainous range in the northern Mexican state of Chihuahua belonging to the Sierra Madre Occidental.\nYour task is to translate from Choguita to English: wiˈʰkâ napaˈwí-ka raˈʔìtʃ͡i-li ˈmí reʔˈlé-ki esˈk extsuperscriptê... | translation | {
"ground_truth": "Many (people) got together to speak over there at the school, but I've never gone.",
"style": "rule"
} | {
"answer": "Many (people) got together to speak over there at the school, but I've never gone.",
"grammar_metadata": [
{
"metadata": {
"Header 1": "",
"Header 2": "",
"Header 3": "",
"Header 4": "",
"source": "/home/zsnajdr/rl-for-low-resource-languages/verl_prepro... | Choguita_English-5-sentences | null | null |
aʔˈlì reˈhòi=ko ˈmá buˈʔí-li aʔˈlì muˈkî=ko ke <> ˈtʃ͡ó | Y el señor ya estaba acostado y la mujer todavía no. | back-to-school | [
{
"content": "Choguita Rarámuri, a Uto-Aztecan language spoken in the Sierra Tarahumara, a mountainous range in the northern Mexican state of Chihuahua belonging to the Sierra Madre Occidental.\nYour task is to translate from Choguita to English: aʔˈlì reˈhòi=ko ˈmá buˈʔí-li aʔˈlì muˈkî=ko ke <> ˈtʃ͡ó\n\nTo he... | translation | {
"ground_truth": "Y el señor ya estaba acostado y la mujer todavía no.",
"style": "rule"
} | {
"answer": "Y el señor ya estaba acostado y la mujer todavía no.",
"grammar_metadata": [
{
"metadata": {
"Header 1": "",
"Header 2": "",
"Header 3": "",
"Header 4": "",
"source": "/home/zsnajdr/rl-for-low-resource-languages/verl_preprocessing/resources//choguita_ra... | Choguita_English-5-sentences | null | null |
ˈkíti ˈbéli=ti ˈwé riˈsó-a=ti ka ru-ˈwá tamuˈhê raˈlàmuli=ko ba | That is why it is said that we are very poor (struggle a lot), us the Rarámuri people. | back-to-school | [{"content":"Choguita Rarámuri, a Uto-Aztecan language spoken in the Sierra Tarahumara, a mountaino(...TRUNCATED) | translation | {"ground_truth":"That is why it is said that we are very poor (struggle a lot), us the Rarámuri peo(...TRUNCATED) | {"answer":"That is why it is said that we are very poor (struggle a lot), us the Rarámuri people.",(...TRUNCATED) | Choguita_English-5-sentences | null | null |
End of preview. Expand in Data Studio
RL-New-Language
Training and evaluation data for the paper. We teach LLMs to translate unseen low-resource languages with RL (GRPO + chrF reward), conditioning the policy on retrieved dictionary entries, parallel sentences, and grammar passages.
Code: https://github.com/hanxuhu/rl-new-language
Split
The main training split (exc3) holds out 6 of 28 translation directions so
they appear only at test time:
- Romansh-Sursilvan ↔ German (similar variety, held-out)
- Romansh-Surmiran ↔ German (similar variety, held-out)
- English → Kalamang (unseen Papuan language)
The remaining 22 directions are used for training.
Files
Training
| File | Use |
|---|---|
train.parquet |
Main RL train set (22 directions, full context: dict + parallel sentences + grammar where available). |
train_nodict.parquet |
Ablation: dictionary removed. |
train_nogram.parquet |
Ablation: grammar removed. |
train_nosent.parquet |
Ablation: parallel sentences removed. |
train_sft.jsonl |
Same data, instruction-tuning JSONL for SFT baselines. |
train_german.parquet |
German ↔ Romansh experiment (appendix). |
Test (paper Table 1)
| File | Paper column |
|---|---|
test_similar.parquet |
Similar — Sursilvan + Surmiran ↔ German (held out from exc3). |
test_indomain_a.parquet |
In-domain — Puter + Vallader ↔ German (with grammar). |
test_indomain_b.parquet |
In-domain — Sutsilvan + Rumantsch ↔ German (no grammar). |
test_kalamang.parquet |
Kalamang — English ↔ Kalamang (held out, unseen language). |
test_unseen4.parquet |
Unseen-4 — English → Dinka, Wolof, Guarani, Kachin. |
Each test file also has variants:
_taskonly— task instruction only (no retrieval context); used for the "no ctx" rows of Table 1._nodict,_nogram,_nosent— matched context drops for the ablation table.
German appendix
test_german_all.parquet plus per-direction files for the German↔Romansh
evaluation.
Format
Each row contains the rendered prompt (prompt), gold translation
(reward_model.ground_truth), and metadata used by the reward function in the
training code.
Reproducing the paper
hf download HanxuHU/rl-new-language --repo-type dataset --local-dir ./data
export HDFS_DATA_PATH=$(pwd)/data
# Then follow the README in https://github.com/hanxuhu/rl-new-language
Data sources
The data is built on top of the following prior work; please cite these papers in addition to ours when using this dataset:
- MTOB / Kalamang test set. Tanzer et al. A Benchmark for Learning to Translate a New Language from One Grammar Book. ICLR 2024.
- Kalamang grammar book. Eline Visser. A Grammar of Kalamang. Language Science Press, 2022. DOI: 10.5281/zenodo.6499927.
- Dinka / Wolof / Guarani / Kachin test sets. Hus & Anastasopoulos. Back to School: Translation Using Grammar Books. EMNLP 2024.
- Romansh parallel data. Provided by Lia Rumantscha (https://www.liarumantscha.ch).
- Downloads last month
- 111