data_source
stringclasses
1 value
prompt
listlengths
2
2
ability
stringclasses
1 value
reward_model
dict
extra_info
dict
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_atc_hard__ce05e186__s000__tb262144__v001\", \"budget\": 2048, \"token_budget\": 262144, \"ppl_base\": 1382.1624478855465, \"nll_base\": 7.231404542922974, \"eval_ref\": \"med_concepts_qa_atc_hard__ce05e186__s000__tb262144__v001\", \"oracle_...
{ "index": "med_concepts_qa_atc_hard__ce05e186__s000__tb262144__v001", "uid": "med_concepts_qa_atc_hard__ce05e186__s000__tb262144__v001", "axis": "domain", "budget": 2048, "token_budget": 262144, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_atc_hard__ce05e186__s000__tb65536__v002\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 1382.1624478855465, \"nll_base\": 7.231404542922974, \"eval_ref\": \"med_concepts_qa_atc_hard__ce05e186__s000__tb65536__v002\", \"oracle_del...
{ "index": "med_concepts_qa_atc_hard__ce05e186__s000__tb65536__v002", "uid": "med_concepts_qa_atc_hard__ce05e186__s000__tb65536__v002", "axis": "domain", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_atc_hard__ce05e186__s001__tb131072__v003\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 1448.5456118109983, \"nll_base\": 7.278315305709839, \"eval_ref\": \"med_concepts_qa_atc_hard__ce05e186__s001__tb131072__v003\", \"oracle_...
{ "index": "med_concepts_qa_atc_hard__ce05e186__s001__tb131072__v003", "uid": "med_concepts_qa_atc_hard__ce05e186__s001__tb131072__v003", "axis": "domain", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_atc_medium__11314e01__s000__tb131072__v000\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 1683.5298931546754, \"nll_base\": 7.428647994995117, \"eval_ref\": \"med_concepts_qa_atc_medium__11314e01__s000__tb131072__v000\", \"ora...
{ "index": "med_concepts_qa_atc_medium__11314e01__s000__tb131072__v000", "uid": "med_concepts_qa_atc_medium__11314e01__s000__tb131072__v000", "axis": "domain", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_atc_medium__11314e01__s000__tb262144__v001\", \"budget\": 2048, \"token_budget\": 262144, \"ppl_base\": 1683.5298931546754, \"nll_base\": 7.428647994995117, \"eval_ref\": \"med_concepts_qa_atc_medium__11314e01__s000__tb262144__v001\", \"ora...
{ "index": "med_concepts_qa_atc_medium__11314e01__s000__tb262144__v001", "uid": "med_concepts_qa_atc_medium__11314e01__s000__tb262144__v001", "axis": "domain", "budget": 2048, "token_budget": 262144, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_atc_medium__11314e01__s000__tb65536__v002\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 1683.5298931546754, \"nll_base\": 7.428647994995117, \"eval_ref\": \"med_concepts_qa_atc_medium__11314e01__s000__tb65536__v002\", \"oracle...
{ "index": "med_concepts_qa_atc_medium__11314e01__s000__tb65536__v002", "uid": "med_concepts_qa_atc_medium__11314e01__s000__tb65536__v002", "axis": "domain", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd10cm_easy__56b2ba2b__s000__tb131072__v002\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 1063.752215004692, \"nll_base\": 6.969557762145996, \"eval_ref\": \"med_concepts_qa_icd10cm_easy__56b2ba2b__s000__tb131072__v002\", \"...
{ "index": "med_concepts_qa_icd10cm_easy__56b2ba2b__s000__tb131072__v002", "uid": "med_concepts_qa_icd10cm_easy__56b2ba2b__s000__tb131072__v002", "axis": "domain", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd10cm_easy__56b2ba2b__s000__tb65536__v001\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 1063.752215004692, \"nll_base\": 6.969557762145996, \"eval_ref\": \"med_concepts_qa_icd10cm_easy__56b2ba2b__s000__tb65536__v001\", \"ora...
{ "index": "med_concepts_qa_icd10cm_easy__56b2ba2b__s000__tb65536__v001", "uid": "med_concepts_qa_icd10cm_easy__56b2ba2b__s000__tb65536__v001", "axis": "domain", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd10cm_hard__6fbd3159__s000__tb131072__v002\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 1131.1295383828847, \"nll_base\": 7.030972003936768, \"eval_ref\": \"med_concepts_qa_icd10cm_hard__6fbd3159__s000__tb131072__v002\", \...
{ "index": "med_concepts_qa_icd10cm_hard__6fbd3159__s000__tb131072__v002", "uid": "med_concepts_qa_icd10cm_hard__6fbd3159__s000__tb131072__v002", "axis": "domain", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd10cm_hard__6fbd3159__s000__tb65536__v001\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 1131.1295383828847, \"nll_base\": 7.030972003936768, \"eval_ref\": \"med_concepts_qa_icd10cm_hard__6fbd3159__s000__tb65536__v001\", \"or...
{ "index": "med_concepts_qa_icd10cm_hard__6fbd3159__s000__tb65536__v001", "uid": "med_concepts_qa_icd10cm_hard__6fbd3159__s000__tb65536__v001", "axis": "domain", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd10proc_easy__4058e328__s000__tb131072__v002\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 2031.0553928363192, \"nll_base\": 7.6163108348846436, \"eval_ref\": \"med_concepts_qa_icd10proc_easy__4058e328__s000__tb131072__v002...
{ "index": "med_concepts_qa_icd10proc_easy__4058e328__s000__tb131072__v002", "uid": "med_concepts_qa_icd10proc_easy__4058e328__s000__tb131072__v002", "axis": "domain", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd10proc_easy__4058e328__s000__tb262144__v000\", \"budget\": 2048, \"token_budget\": 262144, \"ppl_base\": 2031.0553928363192, \"nll_base\": 7.6163108348846436, \"eval_ref\": \"med_concepts_qa_icd10proc_easy__4058e328__s000__tb262144__v000...
{ "index": "med_concepts_qa_icd10proc_easy__4058e328__s000__tb262144__v000", "uid": "med_concepts_qa_icd10proc_easy__4058e328__s000__tb262144__v000", "axis": "domain", "budget": 2048, "token_budget": 262144, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd10proc_easy__4058e328__s000__tb65536__v001\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 2031.0553928363192, \"nll_base\": 7.6163108348846436, \"eval_ref\": \"med_concepts_qa_icd10proc_easy__4058e328__s000__tb65536__v001\",...
{ "index": "med_concepts_qa_icd10proc_easy__4058e328__s000__tb65536__v001", "uid": "med_concepts_qa_icd10proc_easy__4058e328__s000__tb65536__v001", "axis": "domain", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd10proc_easy__4058e328__s001__tb262144__v003\", \"budget\": 2048, \"token_budget\": 262144, \"ppl_base\": 1522.2855268163833, \"nll_base\": 7.327968120574951, \"eval_ref\": \"med_concepts_qa_icd10proc_easy__4058e328__s001__tb262144__v003\...
{ "index": "med_concepts_qa_icd10proc_easy__4058e328__s001__tb262144__v003", "uid": "med_concepts_qa_icd10proc_easy__4058e328__s001__tb262144__v003", "axis": "domain", "budget": 2048, "token_budget": 262144, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd10proc_hard__63f38420__s000__tb131072__v000\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 1901.4411076128745, \"nll_base\": 7.55036735534668, \"eval_ref\": \"med_concepts_qa_icd10proc_hard__63f38420__s000__tb131072__v000\"...
{ "index": "med_concepts_qa_icd10proc_hard__63f38420__s000__tb131072__v000", "uid": "med_concepts_qa_icd10proc_hard__63f38420__s000__tb131072__v000", "axis": "domain", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd10proc_hard__63f38420__s000__tb65536__v002\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 1901.4411076128745, \"nll_base\": 7.55036735534668, \"eval_ref\": \"med_concepts_qa_icd10proc_hard__63f38420__s000__tb65536__v002\", \...
{ "index": "med_concepts_qa_icd10proc_hard__63f38420__s000__tb65536__v002", "uid": "med_concepts_qa_icd10proc_hard__63f38420__s000__tb65536__v002", "axis": "domain", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd10proc_medium__9d288627__s000__tb131072__v001\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 1510.2856049280251, \"nll_base\": 7.320054054260254, \"eval_ref\": \"med_concepts_qa_icd10proc_medium__9d288627__s000__tb131072__v...
{ "index": "med_concepts_qa_icd10proc_medium__9d288627__s000__tb131072__v001", "uid": "med_concepts_qa_icd10proc_medium__9d288627__s000__tb131072__v001", "axis": "domain", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd10proc_medium__9d288627__s000__tb262144__v002\", \"budget\": 2048, \"token_budget\": 262144, \"ppl_base\": 1510.2856049280251, \"nll_base\": 7.320054054260254, \"eval_ref\": \"med_concepts_qa_icd10proc_medium__9d288627__s000__tb262144__v...
{ "index": "med_concepts_qa_icd10proc_medium__9d288627__s000__tb262144__v002", "uid": "med_concepts_qa_icd10proc_medium__9d288627__s000__tb262144__v002", "axis": "domain", "budget": 2048, "token_budget": 262144, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd10proc_medium__9d288627__s000__tb65536__v000\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 1510.2856049280251, \"nll_base\": 7.320054054260254, \"eval_ref\": \"med_concepts_qa_icd10proc_medium__9d288627__s000__tb65536__v000...
{ "index": "med_concepts_qa_icd10proc_medium__9d288627__s000__tb65536__v000", "uid": "med_concepts_qa_icd10proc_medium__9d288627__s000__tb65536__v000", "axis": "domain", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd10proc_medium__9d288627__s001__tb65536__v003\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 1859.3021083772514, \"nll_base\": 7.527956485748291, \"eval_ref\": \"med_concepts_qa_icd10proc_medium__9d288627__s001__tb65536__v003...
{ "index": "med_concepts_qa_icd10proc_medium__9d288627__s001__tb65536__v003", "uid": "med_concepts_qa_icd10proc_medium__9d288627__s001__tb65536__v003", "axis": "domain", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd9cm_easy__f4eedd9e__s000__tb131072__v002\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 1159.11913625668, \"nll_base\": 7.055415630340576, \"eval_ref\": \"med_concepts_qa_icd9cm_easy__f4eedd9e__s000__tb131072__v002\", \"ora...
{ "index": "med_concepts_qa_icd9cm_easy__f4eedd9e__s000__tb131072__v002", "uid": "med_concepts_qa_icd9cm_easy__f4eedd9e__s000__tb131072__v002", "axis": "domain", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd9cm_easy__f4eedd9e__s000__tb262144__v000\", \"budget\": 2048, \"token_budget\": 262144, \"ppl_base\": 1159.11913625668, \"nll_base\": 7.055415630340576, \"eval_ref\": \"med_concepts_qa_icd9cm_easy__f4eedd9e__s000__tb262144__v000\", \"ora...
{ "index": "med_concepts_qa_icd9cm_easy__f4eedd9e__s000__tb262144__v000", "uid": "med_concepts_qa_icd9cm_easy__f4eedd9e__s000__tb262144__v000", "axis": "domain", "budget": 2048, "token_budget": 262144, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd9cm_easy__f4eedd9e__s000__tb65536__v001\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 1159.11913625668, \"nll_base\": 7.055415630340576, \"eval_ref\": \"med_concepts_qa_icd9cm_easy__f4eedd9e__s000__tb65536__v001\", \"oracle...
{ "index": "med_concepts_qa_icd9cm_easy__f4eedd9e__s000__tb65536__v001", "uid": "med_concepts_qa_icd9cm_easy__f4eedd9e__s000__tb65536__v001", "axis": "domain", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd9cm_medium__8dccd159__s000__tb131072__v001\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 816.3549459464714, \"nll_base\": 6.7048492431640625, \"eval_ref\": \"med_concepts_qa_icd9cm_medium__8dccd159__s000__tb131072__v001\",...
{ "index": "med_concepts_qa_icd9cm_medium__8dccd159__s000__tb131072__v001", "uid": "med_concepts_qa_icd9cm_medium__8dccd159__s000__tb131072__v001", "axis": "domain", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd9cm_medium__8dccd159__s000__tb262144__v002\", \"budget\": 2048, \"token_budget\": 262144, \"ppl_base\": 816.3549459464714, \"nll_base\": 6.7048492431640625, \"eval_ref\": \"med_concepts_qa_icd9cm_medium__8dccd159__s000__tb262144__v002\",...
{ "index": "med_concepts_qa_icd9cm_medium__8dccd159__s000__tb262144__v002", "uid": "med_concepts_qa_icd9cm_medium__8dccd159__s000__tb262144__v002", "axis": "domain", "budget": 2048, "token_budget": 262144, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd9cm_medium__8dccd159__s000__tb65536__v000\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 816.3549459464714, \"nll_base\": 6.7048492431640625, \"eval_ref\": \"med_concepts_qa_icd9cm_medium__8dccd159__s000__tb65536__v000\", \"...
{ "index": "med_concepts_qa_icd9cm_medium__8dccd159__s000__tb65536__v000", "uid": "med_concepts_qa_icd9cm_medium__8dccd159__s000__tb65536__v000", "axis": "domain", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd9cm_medium__8dccd159__s001__tb65536__v003\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 955.869185994098, \"nll_base\": 6.862621068954468, \"eval_ref\": \"med_concepts_qa_icd9cm_medium__8dccd159__s001__tb65536__v003\", \"or...
{ "index": "med_concepts_qa_icd9cm_medium__8dccd159__s001__tb65536__v003", "uid": "med_concepts_qa_icd9cm_medium__8dccd159__s001__tb65536__v003", "axis": "domain", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd9proc_easy__34ef6201__s000__tb131072__v000\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 1128.0079661089608, \"nll_base\": 7.028208494186401, \"eval_ref\": \"med_concepts_qa_icd9proc_easy__34ef6201__s000__tb131072__v000\",...
{ "index": "med_concepts_qa_icd9proc_easy__34ef6201__s000__tb131072__v000", "uid": "med_concepts_qa_icd9proc_easy__34ef6201__s000__tb131072__v000", "axis": "domain", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd9proc_hard__b3984895__s000__tb131072__v000\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 1250.96492813016, \"nll_base\": 7.1316704750061035, \"eval_ref\": \"med_concepts_qa_icd9proc_hard__b3984895__s000__tb131072__v000\", ...
{ "index": "med_concepts_qa_icd9proc_hard__b3984895__s000__tb131072__v000", "uid": "med_concepts_qa_icd9proc_hard__b3984895__s000__tb131072__v000", "axis": "domain", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd9proc_hard__b3984895__s000__tb65536__v002\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 1250.96492813016, \"nll_base\": 7.1316704750061035, \"eval_ref\": \"med_concepts_qa_icd9proc_hard__b3984895__s000__tb65536__v002\", \"o...
{ "index": "med_concepts_qa_icd9proc_hard__b3984895__s000__tb65536__v002", "uid": "med_concepts_qa_icd9proc_hard__b3984895__s000__tb65536__v002", "axis": "domain", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd9proc_medium__dfe1372d__s000__tb131072__v000\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 914.7257174987559, \"nll_base\": 6.818624258041382, \"eval_ref\": \"med_concepts_qa_icd9proc_medium__dfe1372d__s000__tb131072__v000...
{ "index": "med_concepts_qa_icd9proc_medium__dfe1372d__s000__tb131072__v000", "uid": "med_concepts_qa_icd9proc_medium__dfe1372d__s000__tb131072__v000", "axis": "domain", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd9proc_medium__dfe1372d__s000__tb65536__v002\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 914.7257174987559, \"nll_base\": 6.818624258041382, \"eval_ref\": \"med_concepts_qa_icd9proc_medium__dfe1372d__s000__tb65536__v002\",...
{ "index": "med_concepts_qa_icd9proc_medium__dfe1372d__s000__tb65536__v002", "uid": "med_concepts_qa_icd9proc_medium__dfe1372d__s000__tb65536__v002", "axis": "domain", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_concepts_qa_icd9proc_medium__dfe1372d__s001__tb131072__v003\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 1242.254189877015, \"nll_base\": 7.124682903289795, \"eval_ref\": \"med_concepts_qa_icd9proc_medium__dfe1372d__s001__tb131072__v003...
{ "index": "med_concepts_qa_icd9proc_medium__dfe1372d__s001__tb131072__v003", "uid": "med_concepts_qa_icd9proc_medium__dfe1372d__s001__tb131072__v003", "axis": "domain", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_prescriptions_hard__d49a4101__s000__tb131072__v000\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 3.4617092617211, \"nll_base\": 1.2417624733664774, \"eval_ref\": \"med_prescriptions_hard__d49a4101__s000__tb131072__v000\", \"oracle_delta\...
{ "index": "med_prescriptions_hard__d49a4101__s000__tb131072__v000", "uid": "med_prescriptions_hard__d49a4101__s000__tb131072__v000", "axis": "domain", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_prescriptions_hard__d49a4101__s000__tb262144__v001\", \"budget\": 2048, \"token_budget\": 262144, \"ppl_base\": 3.4874944191229806, \"nll_base\": 1.2491835468947285, \"eval_ref\": \"med_prescriptions_hard__d49a4101__s000__tb262144__v001\", \"oracle_del...
{ "index": "med_prescriptions_hard__d49a4101__s000__tb262144__v001", "uid": "med_prescriptions_hard__d49a4101__s000__tb262144__v001", "axis": "domain", "budget": 2048, "token_budget": 262144, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_prescriptions_hard__d49a4101__s000__tb65536__v002\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 3.461071407613554, \"nll_base\": 1.2415781965159407, \"eval_ref\": \"med_prescriptions_hard__d49a4101__s000__tb65536__v002\", \"oracle_delta\"...
{ "index": "med_prescriptions_hard__d49a4101__s000__tb65536__v002", "uid": "med_prescriptions_hard__d49a4101__s000__tb65536__v002", "axis": "domain", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"med_prescriptions_hard__d49a4101__s001__tb131072__v003\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 4.1495255470209, \"nll_base\": 1.4229940016878293, \"eval_ref\": \"med_prescriptions_hard__d49a4101__s001__tb131072__v003\", \"oracle_delta\...
{ "index": "med_prescriptions_hard__d49a4101__s001__tb131072__v003", "uid": "med_prescriptions_hard__d49a4101__s001__tb131072__v003", "axis": "domain", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"medqa_4options__79800928__s000__tb131072__v000\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 2568.329070730162, \"nll_base\": 7.851010799407959, \"eval_ref\": \"medqa_4options__79800928__s000__tb131072__v000\", \"oracle_delta\": 0.7117137908...
{ "index": "medqa_4options__79800928__s000__tb131072__v000", "uid": "medqa_4options__79800928__s000__tb131072__v000", "axis": "domain", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"medqa_4options__79800928__s000__tb65536__v002\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 2568.329070730162, \"nll_base\": 7.851010799407959, \"eval_ref\": \"medqa_4options__79800928__s000__tb65536__v002\", \"oracle_delta\": 0.5950314998626...
{ "index": "medqa_4options__79800928__s000__tb65536__v002", "uid": "medqa_4options__79800928__s000__tb65536__v002", "axis": "domain", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"metabench_arc__bb35525c__s000__tb131072__v000\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 11.840356814389889, \"nll_base\": 2.4715137653522663, \"eval_ref\": \"metabench_arc__bb35525c__s000__tb131072__v000\", \"oracle_delta\": 0.1341926643...
{ "index": "metabench_arc__bb35525c__s000__tb131072__v000", "uid": "metabench_arc__bb35525c__s000__tb131072__v000", "axis": "instruction", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"metabench_arc_secondary__8af1c93e__s000__tb65536__v000\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 9.247281630350582, \"nll_base\": 2.224329630533854, \"eval_ref\": \"metabench_arc_secondary__8af1c93e__s000__tb65536__v000\", \"oracle_delta\...
{ "index": "metabench_arc_secondary__8af1c93e__s000__tb65536__v000", "uid": "metabench_arc_secondary__8af1c93e__s000__tb65536__v000", "axis": "instruction", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"metabench_hellaswag__1784b5fc__s000__tb131072__v000\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 1.8086556061102206, \"nll_base\": 0.5925838102404334, \"eval_ref\": \"metabench_hellaswag__1784b5fc__s000__tb131072__v000\", \"oracle_delta\": ...
{ "index": "metabench_hellaswag__1784b5fc__s000__tb131072__v000", "uid": "metabench_hellaswag__1784b5fc__s000__tb131072__v000", "axis": "instruction", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"metabench_hellaswag_permute__14b19ecd__s000__tb262144__v000\", \"budget\": 2048, \"token_budget\": 262144, \"ppl_base\": 1.9422386712524369, \"nll_base\": 0.6638412620356232, \"eval_ref\": \"metabench_hellaswag_permute__14b19ecd__s000__tb262144__v000\", \"...
{ "index": "metabench_hellaswag_permute__14b19ecd__s000__tb262144__v000", "uid": "metabench_hellaswag_permute__14b19ecd__s000__tb262144__v000", "axis": "instruction", "budget": 2048, "token_budget": 262144, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"metabench_hellaswag_secondary__c1c76adc__s000__tb131072__v000\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 1.9990342805522696, \"nll_base\": 0.6926642042217832, \"eval_ref\": \"metabench_hellaswag_secondary__c1c76adc__s000__tb131072__v000\"...
{ "index": "metabench_hellaswag_secondary__c1c76adc__s000__tb131072__v000", "uid": "metabench_hellaswag_secondary__c1c76adc__s000__tb131072__v000", "axis": "instruction", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"metabench_mmlu__bd61f983__s000__tb262144__v000\", \"budget\": 2048, \"token_budget\": 262144, \"ppl_base\": 1373.53528383221, \"nll_base\": 7.225143194198608, \"eval_ref\": \"metabench_mmlu__bd61f983__s000__tb262144__v000\", \"oracle_delta\": 1.26808834075...
{ "index": "metabench_mmlu__bd61f983__s000__tb262144__v000", "uid": "metabench_mmlu__bd61f983__s000__tb262144__v000", "axis": "instruction", "budget": 2048, "token_budget": 262144, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"metabench_winogrande_secondary_permute__b0897462__s000__tb131072__v000\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 92.81495280660161, \"nll_base\": 4.530607756193693, \"eval_ref\": \"metabench_winogrande_secondary_permute__b0897462__s000__...
{ "index": "metabench_winogrande_secondary_permute__b0897462__s000__tb131072__v000", "uid": "metabench_winogrande_secondary_permute__b0897462__s000__tb131072__v000", "axis": "instruction", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mgsm_direct_ca__a9190cb4__s000__tb65536__v000\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 330.60577625963674, \"nll_base\": 5.800926657284007, \"eval_ref\": \"mgsm_direct_ca__a9190cb4__s000__tb65536__v000\", \"oracle_delta\": 0.368868958716...
{ "index": "mgsm_direct_ca__a9190cb4__s000__tb65536__v000", "uid": "mgsm_direct_ca__a9190cb4__s000__tb65536__v000", "axis": "multilingual", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mgsm_direct_de__ec7bf57d__s000__tb262144__v000\", \"budget\": 2048, \"token_budget\": 262144, \"ppl_base\": 352.32546288811074, \"nll_base\": 5.864555358886719, \"eval_ref\": \"mgsm_direct_de__ec7bf57d__s000__tb262144__v000\", \"oracle_delta\": 0.746277618...
{ "index": "mgsm_direct_de__ec7bf57d__s000__tb262144__v000", "uid": "mgsm_direct_de__ec7bf57d__s000__tb262144__v000", "axis": "math", "budget": 2048, "token_budget": 262144, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mgsm_direct_de__ec7bf57d__s000__tb65536__v001\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 352.32546288811074, \"nll_base\": 5.864555358886719, \"eval_ref\": \"mgsm_direct_de__ec7bf57d__s000__tb65536__v001\", \"oracle_delta\": 0.075439147949...
{ "index": "mgsm_direct_de__ec7bf57d__s000__tb65536__v001", "uid": "mgsm_direct_de__ec7bf57d__s000__tb65536__v001", "axis": "math", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mgsm_direct_en__c79b5f79__s000__tb131072__v000\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 205.75765063083622, \"nll_base\": 5.326699022976857, \"eval_ref\": \"mgsm_direct_en__c79b5f79__s000__tb131072__v000\", \"oracle_delta\": 0.521264706...
{ "index": "mgsm_direct_en__c79b5f79__s000__tb131072__v000", "uid": "mgsm_direct_en__c79b5f79__s000__tb131072__v000", "axis": "math", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mgsm_direct_ja__759864d7__s000__tb65536__v001\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 269.9386271051873, \"nll_base\": 5.598194626142394, \"eval_ref\": \"mgsm_direct_ja__759864d7__s000__tb65536__v001\", \"oracle_delta\": 0.1523938448923...
{ "index": "mgsm_direct_ja__759864d7__s000__tb65536__v001", "uid": "mgsm_direct_ja__759864d7__s000__tb65536__v001", "axis": "math", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mgsm_direct_zh__39176631__s000__tb65536__v001\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 638.3300438037127, \"nll_base\": 6.458855459623248, \"eval_ref\": \"mgsm_direct_zh__39176631__s000__tb65536__v001\", \"oracle_delta\": 0.1465405615690...
{ "index": "mgsm_direct_zh__39176631__s000__tb65536__v001", "uid": "mgsm_direct_zh__39176631__s000__tb65536__v001", "axis": "math", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mgsm_en_cot_de__7ab52e2d__s000__tb262144__v000\", \"budget\": 2048, \"token_budget\": 262144, \"ppl_base\": 611.7851870966148, \"nll_base\": 6.41638121942077, \"eval_ref\": \"mgsm_en_cot_de__7ab52e2d__s000__tb262144__v000\", \"oracle_delta\": 1.12330735331...
{ "index": "mgsm_en_cot_de__7ab52e2d__s000__tb262144__v000", "uid": "mgsm_en_cot_de__7ab52e2d__s000__tb262144__v000", "axis": "math", "budget": 2048, "token_budget": 262144, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mgsm_en_cot_de__7ab52e2d__s000__tb65536__v001\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 611.7851870966148, \"nll_base\": 6.41638121942077, \"eval_ref\": \"mgsm_en_cot_de__7ab52e2d__s000__tb65536__v001\", \"oracle_delta\": 0.27156236436631...
{ "index": "mgsm_en_cot_de__7ab52e2d__s000__tb65536__v001", "uid": "mgsm_en_cot_de__7ab52e2d__s000__tb65536__v001", "axis": "math", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mgsm_en_cot_en__c2258a2e__s000__tb131072__v001\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 161.55009394058484, \"nll_base\": 5.084815273765757, \"eval_ref\": \"mgsm_en_cot_en__c2258a2e__s000__tb131072__v001\", \"oracle_delta\": 0.362616274...
{ "index": "mgsm_en_cot_en__c2258a2e__s000__tb131072__v001", "uid": "mgsm_en_cot_en__c2258a2e__s000__tb131072__v001", "axis": "math", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mgsm_en_cot_en__c2258a2e__s000__tb65536__v000\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 161.55009394058484, \"nll_base\": 5.084815273765757, \"eval_ref\": \"mgsm_en_cot_en__c2258a2e__s000__tb65536__v000\", \"oracle_delta\": 0.138879343241...
{ "index": "mgsm_en_cot_en__c2258a2e__s000__tb65536__v000", "uid": "mgsm_en_cot_en__c2258a2e__s000__tb65536__v000", "axis": "math", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mgsm_en_cot_zh__1db46641__s000__tb131072__v001\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 734.1593051278915, \"nll_base\": 6.598726042024382, \"eval_ref\": \"mgsm_en_cot_zh__1db46641__s000__tb131072__v001\", \"oracle_delta\": 0.3132971459...
{ "index": "mgsm_en_cot_zh__1db46641__s000__tb131072__v001", "uid": "mgsm_en_cot_zh__1db46641__s000__tb131072__v001", "axis": "math", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mgsm_en_cot_zh__1db46641__s000__tb65536__v000\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 734.1593051278915, \"nll_base\": 6.598726042024382, \"eval_ref\": \"mgsm_en_cot_zh__1db46641__s000__tb65536__v000\", \"oracle_delta\": 0.1845552213899...
{ "index": "mgsm_en_cot_zh__1db46641__s000__tb65536__v000", "uid": "mgsm_en_cot_zh__1db46641__s000__tb65536__v000", "axis": "math", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mgsm_native_cot_de__4fc7a02f__s000__tb262144__v000\", \"budget\": 2048, \"token_budget\": 262144, \"ppl_base\": 391.8350257514747, \"nll_base\": 5.970840898532312, \"eval_ref\": \"mgsm_native_cot_de__4fc7a02f__s000__tb262144__v000\", \"oracle_delta\": 0.69...
{ "index": "mgsm_native_cot_de__4fc7a02f__s000__tb262144__v000", "uid": "mgsm_native_cot_de__4fc7a02f__s000__tb262144__v000", "axis": "math", "budget": 2048, "token_budget": 262144, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mgsm_native_cot_de__4fc7a02f__s000__tb65536__v001\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 391.8350257514747, \"nll_base\": 5.970840898532312, \"eval_ref\": \"mgsm_native_cot_de__4fc7a02f__s000__tb65536__v001\", \"oracle_delta\": 0.18908...
{ "index": "mgsm_native_cot_de__4fc7a02f__s000__tb65536__v001", "uid": "mgsm_native_cot_de__4fc7a02f__s000__tb65536__v001", "axis": "math", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mgsm_native_cot_en__83ca51eb__s000__tb65536__v001\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 238.33032830387967, \"nll_base\": 5.47365764543122, \"eval_ref\": \"mgsm_native_cot_en__83ca51eb__s000__tb65536__v001\", \"oracle_delta\": 0.17932...
{ "index": "mgsm_native_cot_en__83ca51eb__s000__tb65536__v001", "uid": "mgsm_native_cot_en__83ca51eb__s000__tb65536__v001", "axis": "math", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mgsm_native_cot_ja__0006ff82__s000__tb131072__v000\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 250.67044634029577, \"nll_base\": 5.524139113653273, \"eval_ref\": \"mgsm_native_cot_ja__0006ff82__s000__tb131072__v000\", \"oracle_delta\": 0.2...
{ "index": "mgsm_native_cot_ja__0006ff82__s000__tb131072__v000", "uid": "mgsm_native_cot_ja__0006ff82__s000__tb131072__v000", "axis": "math", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mgsm_native_cot_zh__71b9d147__s000__tb131072__v000\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 301.8187621453214, \"nll_base\": 5.709826711880958, \"eval_ref\": \"mgsm_native_cot_zh__71b9d147__s000__tb131072__v000\", \"oracle_delta\": 0.19...
{ "index": "mgsm_native_cot_zh__71b9d147__s000__tb131072__v000", "uid": "mgsm_native_cot_zh__71b9d147__s000__tb131072__v000", "axis": "math", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"minerva_math500__33bce439__s000__tb262144__v000\", \"budget\": 2048, \"token_budget\": 262144, \"ppl_base\": 72.01332051779855, \"nll_base\": 4.276851109095982, \"eval_ref\": \"minerva_math500__33bce439__s000__tb262144__v000\", \"oracle_delta\": 0.25113172...
{ "index": "minerva_math500__33bce439__s000__tb262144__v000", "uid": "minerva_math500__33bce439__s000__tb262144__v000", "axis": "math", "budget": 2048, "token_budget": 262144, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"minerva_math500__33bce439__s000__tb65536__v001\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 72.01332051779855, \"nll_base\": 4.276851109095982, \"eval_ref\": \"minerva_math500__33bce439__s000__tb65536__v001\", \"oracle_delta\": 0.05020973899...
{ "index": "minerva_math500__33bce439__s000__tb65536__v001", "uid": "minerva_math500__33bce439__s000__tb65536__v001", "axis": "math", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"minerva_math_algebra__f4c0c5cd__s000__tb131072__v001\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 172.64826035633035, \"nll_base\": 5.15125634765625, \"eval_ref\": \"minerva_math_algebra__f4c0c5cd__s000__tb131072__v001\", \"oracle_delta\": ...
{ "index": "minerva_math_algebra__f4c0c5cd__s000__tb131072__v001", "uid": "minerva_math_algebra__f4c0c5cd__s000__tb131072__v001", "axis": "math", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"minerva_math_geometry__d0e24525__s000__tb131072__v000\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 70.86824110873428, \"nll_base\": 4.260822393955329, \"eval_ref\": \"minerva_math_geometry__d0e24525__s000__tb131072__v000\", \"oracle_delta\"...
{ "index": "minerva_math_geometry__d0e24525__s000__tb131072__v000", "uid": "minerva_math_geometry__d0e24525__s000__tb131072__v000", "axis": "math", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"minerva_math_geometry__d0e24525__s000__tb262144__v001\", \"budget\": 2048, \"token_budget\": 262144, \"ppl_base\": 70.86824110873428, \"nll_base\": 4.260822393955329, \"eval_ref\": \"minerva_math_geometry__d0e24525__s000__tb262144__v001\", \"oracle_delta\"...
{ "index": "minerva_math_geometry__d0e24525__s000__tb262144__v001", "uid": "minerva_math_geometry__d0e24525__s000__tb262144__v001", "axis": "math", "budget": 2048, "token_budget": 262144, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"minerva_math_intermediate_algebra__9482de00__s000__tb131072__v001\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 61.02982116189908, \"nll_base\": 4.11136261622111, \"eval_ref\": \"minerva_math_intermediate_algebra__9482de00__s000__tb131072__v...
{ "index": "minerva_math_intermediate_algebra__9482de00__s000__tb131072__v001", "uid": "minerva_math_intermediate_algebra__9482de00__s000__tb131072__v001", "axis": "math", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"minerva_math_intermediate_algebra__9482de00__s000__tb65536__v000\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 61.02982116189908, \"nll_base\": 4.11136261622111, \"eval_ref\": \"minerva_math_intermediate_algebra__9482de00__s000__tb65536__v000...
{ "index": "minerva_math_intermediate_algebra__9482de00__s000__tb65536__v000", "uid": "minerva_math_intermediate_algebra__9482de00__s000__tb65536__v000", "axis": "math", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"minerva_math_num_theory__a12df733__s000__tb65536__v001\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 131.1256754841931, \"nll_base\": 4.876156218155571, \"eval_ref\": \"minerva_math_num_theory__a12df733__s000__tb65536__v001\", \"oracle_delta\...
{ "index": "minerva_math_num_theory__a12df733__s000__tb65536__v001", "uid": "minerva_math_num_theory__a12df733__s000__tb65536__v001", "axis": "math", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"minerva_math_precalc__cf580ef0__s000__tb131072__v000\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 24.568857235854587, \"nll_base\": 3.20147967490421, \"eval_ref\": \"minerva_math_precalc__cf580ef0__s000__tb131072__v000\", \"oracle_delta\": ...
{ "index": "minerva_math_precalc__cf580ef0__s000__tb131072__v000", "uid": "minerva_math_precalc__cf580ef0__s000__tb131072__v000", "axis": "math", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"minerva_math_precalc__cf580ef0__s000__tb262144__v001\", \"budget\": 2048, \"token_budget\": 262144, \"ppl_base\": 24.568857235854587, \"nll_base\": 3.20147967490421, \"eval_ref\": \"minerva_math_precalc__cf580ef0__s000__tb262144__v001\", \"oracle_delta\": ...
{ "index": "minerva_math_precalc__cf580ef0__s000__tb262144__v001", "uid": "minerva_math_precalc__cf580ef0__s000__tb262144__v001", "axis": "math", "budget": 2048, "token_budget": 262144, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_ar_ar__86c93b0c__s002__tb65536__v008\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 3.2296456445507498, \"nll_base\": 1.1723724236468223, \"eval_ref\": \"mlqa_ar_ar__86c93b0c__s002__tb65536__v008\", \"oracle_delta\": 0.024287099073707852,...
{ "index": "mlqa_ar_ar__86c93b0c__s002__tb65536__v008", "uid": "mlqa_ar_ar__86c93b0c__s002__tb65536__v008", "axis": "qa", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_ar_de__62f6ffc3__s000__tb65536__v002\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 6.222013077236026, \"nll_base\": 1.828093500246947, \"eval_ref\": \"mlqa_ar_de__62f6ffc3__s000__tb65536__v002\", \"oracle_delta\": 0.03674926757812513, \"...
{ "index": "mlqa_ar_de__62f6ffc3__s000__tb65536__v002", "uid": "mlqa_ar_de__62f6ffc3__s000__tb65536__v002", "axis": "qa", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_ar_de__62f6ffc3__s001__tb65536__v005\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 6.6361502421206335, \"nll_base\": 1.8925320125361371, \"eval_ref\": \"mlqa_ar_de__62f6ffc3__s001__tb65536__v005\", \"oracle_delta\": 0.03365611609908448, ...
{ "index": "mlqa_ar_de__62f6ffc3__s001__tb65536__v005", "uid": "mlqa_ar_de__62f6ffc3__s001__tb65536__v005", "axis": "qa", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_ar_en__d692bb11__s000__tb65536__v000\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 7.167908985266442, \"nll_base\": 1.9696139782240543, \"eval_ref\": \"mlqa_ar_en__d692bb11__s000__tb65536__v000\", \"oracle_delta\": 0.02498821734406964, \...
{ "index": "mlqa_ar_en__d692bb11__s000__tb65536__v000", "uid": "mlqa_ar_en__d692bb11__s000__tb65536__v000", "axis": "qa", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_ar_en__d692bb11__s002__tb65536__v006\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 4.204303108965118, \"nll_base\": 1.4361085505500388, \"eval_ref\": \"mlqa_ar_en__d692bb11__s002__tb65536__v006\", \"oracle_delta\": 0.029945967724761324, ...
{ "index": "mlqa_ar_en__d692bb11__s002__tb65536__v006", "uid": "mlqa_ar_en__d692bb11__s002__tb65536__v006", "axis": "qa", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_ar_zh__748a66bb__s000__tb65536__v000\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 5.284163762287033, \"nll_base\": 1.6647143782856308, \"eval_ref\": \"mlqa_ar_zh__748a66bb__s000__tb65536__v000\", \"oracle_delta\": 0.03665896977219618, \...
{ "index": "mlqa_ar_zh__748a66bb__s000__tb65536__v000", "uid": "mlqa_ar_zh__748a66bb__s000__tb65536__v000", "axis": "qa", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_de_ar__6ec7a94d__s000__tb65536__v000\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 9.070361614926634, \"nll_base\": 2.205012132673699, \"eval_ref\": \"mlqa_de_ar__6ec7a94d__s000__tb65536__v000\", \"oracle_delta\": 0.03179404941307107, \"...
{ "index": "mlqa_de_ar__6ec7a94d__s000__tb65536__v000", "uid": "mlqa_de_ar__6ec7a94d__s000__tb65536__v000", "axis": "qa", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_de_ar__6ec7a94d__s001__tb131072__v004\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 24.952175952646414, \"nll_base\": 3.2169610309256544, \"eval_ref\": \"mlqa_de_ar__6ec7a94d__s001__tb131072__v004\", \"oracle_delta\": 0.2548259638707133...
{ "index": "mlqa_de_ar__6ec7a94d__s001__tb131072__v004", "uid": "mlqa_de_ar__6ec7a94d__s001__tb131072__v004", "axis": "qa", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_de_ar__6ec7a94d__s001__tb262144__v005\", \"budget\": 2048, \"token_budget\": 262144, \"ppl_base\": 24.952175952646414, \"nll_base\": 3.2169610309256544, \"eval_ref\": \"mlqa_de_ar__6ec7a94d__s001__tb262144__v005\", \"oracle_delta\": 0.5924841264525047...
{ "index": "mlqa_de_ar__6ec7a94d__s001__tb262144__v005", "uid": "mlqa_de_ar__6ec7a94d__s001__tb262144__v005", "axis": "qa", "budget": 2048, "token_budget": 262144, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_de_ar__6ec7a94d__s001__tb65536__v003\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 24.952175952646414, \"nll_base\": 3.2169610309256544, \"eval_ref\": \"mlqa_de_ar__6ec7a94d__s001__tb65536__v003\", \"oracle_delta\": 0.09146911400750257, ...
{ "index": "mlqa_de_ar__6ec7a94d__s001__tb65536__v003", "uid": "mlqa_de_ar__6ec7a94d__s001__tb65536__v003", "axis": "qa", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_de_ar__6ec7a94d__s002__tb131072__v007\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 21.414844953041463, \"nll_base\": 3.0640843709309897, \"eval_ref\": \"mlqa_de_ar__6ec7a94d__s002__tb131072__v007\", \"oracle_delta\": 0.5316280018199575...
{ "index": "mlqa_de_ar__6ec7a94d__s002__tb131072__v007", "uid": "mlqa_de_ar__6ec7a94d__s002__tb131072__v007", "axis": "qa", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_de_ar__6ec7a94d__s002__tb65536__v006\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 21.414844953041463, \"nll_base\": 3.0640843709309897, \"eval_ref\": \"mlqa_de_ar__6ec7a94d__s002__tb65536__v006\", \"oracle_delta\": 0.23310401222922605, ...
{ "index": "mlqa_de_ar__6ec7a94d__s002__tb65536__v006", "uid": "mlqa_de_ar__6ec7a94d__s002__tb65536__v006", "axis": "qa", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_de_de__9b5d8493__s000__tb131072__v002\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 16.98987830250628, \"nll_base\": 2.8326177727686215, \"eval_ref\": \"mlqa_de_de__9b5d8493__s000__tb131072__v002\", \"oracle_delta\": 0.24187594897126496...
{ "index": "mlqa_de_de__9b5d8493__s000__tb131072__v002", "uid": "mlqa_de_de__9b5d8493__s000__tb131072__v002", "axis": "qa", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_de_de__9b5d8493__s000__tb65536__v001\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 16.98987830250628, \"nll_base\": 2.8326177727686215, \"eval_ref\": \"mlqa_de_de__9b5d8493__s000__tb65536__v001\", \"oracle_delta\": 0.13316888678563776, \...
{ "index": "mlqa_de_de__9b5d8493__s000__tb65536__v001", "uid": "mlqa_de_de__9b5d8493__s000__tb65536__v001", "axis": "qa", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_de_de__9b5d8493__s001__tb131072__v005\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 9.912672553103608, \"nll_base\": 2.293813994436553, \"eval_ref\": \"mlqa_de_de__9b5d8493__s001__tb131072__v005\", \"oracle_delta\": 0.2446343624230587, ...
{ "index": "mlqa_de_de__9b5d8493__s001__tb131072__v005", "uid": "mlqa_de_de__9b5d8493__s001__tb131072__v005", "axis": "qa", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_de_de__9b5d8493__s001__tb65536__v004\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 9.912672553103608, \"nll_base\": 2.293813994436553, \"eval_ref\": \"mlqa_de_de__9b5d8493__s001__tb65536__v004\", \"oracle_delta\": 0.0998344652580494, \"r...
{ "index": "mlqa_de_de__9b5d8493__s001__tb65536__v004", "uid": "mlqa_de_de__9b5d8493__s001__tb65536__v004", "axis": "qa", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_de_de__9b5d8493__s002__tb65536__v007\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 18.575627456773887, \"nll_base\": 2.921850369671198, \"eval_ref\": \"mlqa_de_de__9b5d8493__s002__tb65536__v007\", \"oracle_delta\": 0.10926206093134816, \...
{ "index": "mlqa_de_de__9b5d8493__s002__tb65536__v007", "uid": "mlqa_de_de__9b5d8493__s002__tb65536__v007", "axis": "qa", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_de_es__07e2912c__s000__tb131072__v000\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 12.413960875070542, \"nll_base\": 2.5188217163085938, \"eval_ref\": \"mlqa_de_es__07e2912c__s000__tb131072__v000\", \"oracle_delta\": 0.1850881373628659...
{ "index": "mlqa_de_es__07e2912c__s000__tb131072__v000", "uid": "mlqa_de_es__07e2912c__s000__tb131072__v000", "axis": "qa", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_de_es__07e2912c__s000__tb262144__v001\", \"budget\": 2048, \"token_budget\": 262144, \"ppl_base\": 12.413960875070542, \"nll_base\": 2.5188217163085938, \"eval_ref\": \"mlqa_de_es__07e2912c__s000__tb262144__v001\", \"oracle_delta\": 0.3175024478993515...
{ "index": "mlqa_de_es__07e2912c__s000__tb262144__v001", "uid": "mlqa_de_es__07e2912c__s000__tb262144__v001", "axis": "qa", "budget": 2048, "token_budget": 262144, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_de_es__07e2912c__s000__tb65536__v002\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 12.413960875070542, \"nll_base\": 2.5188217163085938, \"eval_ref\": \"mlqa_de_es__07e2912c__s000__tb65536__v002\", \"oracle_delta\": 0.07520488982504991, ...
{ "index": "mlqa_de_es__07e2912c__s000__tb65536__v002", "uid": "mlqa_de_es__07e2912c__s000__tb65536__v002", "axis": "qa", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_de_es__07e2912c__s001__tb131072__v003\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 8.444539339250413, \"nll_base\": 2.133520000429783, \"eval_ref\": \"mlqa_de_es__07e2912c__s001__tb131072__v003\", \"oracle_delta\": 0.13015955992607742,...
{ "index": "mlqa_de_es__07e2912c__s001__tb131072__v003", "uid": "mlqa_de_es__07e2912c__s001__tb131072__v003", "axis": "qa", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_de_es__07e2912c__s001__tb65536__v005\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 8.444539339250413, \"nll_base\": 2.133520000429783, \"eval_ref\": \"mlqa_de_es__07e2912c__s001__tb65536__v005\", \"oracle_delta\": 0.03833224196305807, \"...
{ "index": "mlqa_de_es__07e2912c__s001__tb65536__v005", "uid": "mlqa_de_es__07e2912c__s001__tb65536__v005", "axis": "qa", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_de_es__07e2912c__s002__tb131072__v006\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 15.236492731641793, \"nll_base\": 2.7236933880541696, \"eval_ref\": \"mlqa_de_es__07e2912c__s002__tb131072__v006\", \"oracle_delta\": 0.1671258397849206...
{ "index": "mlqa_de_es__07e2912c__s002__tb131072__v006", "uid": "mlqa_de_es__07e2912c__s002__tb131072__v006", "axis": "qa", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_de_hi__36474374__s000__tb131072__v002\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 13.07476721625129, \"nll_base\": 2.5706842060641075, \"eval_ref\": \"mlqa_de_hi__36474374__s000__tb131072__v002\", \"oracle_delta\": 0.305555926258541, ...
{ "index": "mlqa_de_hi__36474374__s000__tb131072__v002", "uid": "mlqa_de_hi__36474374__s000__tb131072__v002", "axis": "qa", "budget": 2048, "token_budget": 131072, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_de_hi__36474374__s000__tb262144__v000\", \"budget\": 2048, \"token_budget\": 262144, \"ppl_base\": 13.07476721625129, \"nll_base\": 2.5706842060641075, \"eval_ref\": \"mlqa_de_hi__36474374__s000__tb262144__v000\", \"oracle_delta\": 0.5009978561156045,...
{ "index": "mlqa_de_hi__36474374__s000__tb262144__v000", "uid": "mlqa_de_hi__36474374__s000__tb262144__v000", "axis": "qa", "budget": 2048, "token_budget": 262144, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_de_hi__36474374__s000__tb65536__v001\", \"budget\": 2048, \"token_budget\": 65536, \"ppl_base\": 13.07476721625129, \"nll_base\": 2.5706842060641075, \"eval_ref\": \"mlqa_de_hi__36474374__s000__tb65536__v001\", \"oracle_delta\": 0.16041403031425805, \...
{ "index": "mlqa_de_hi__36474374__s000__tb65536__v001", "uid": "mlqa_de_hi__36474374__s000__tb65536__v001", "axis": "qa", "budget": 2048, "token_budget": 65536, "variant": "filter" }
datacuration
[ { "role": "system", "content": "You are a **training-data curator**. A small base language model needs to do better on a **target task**, and you must choose which examples from a large pool of candidate training examples it should be fine-tuned on.\n\nThe base model will be LoRA-fine-tuned on exactly the e...
datacuration_selection
{ "style": "datacuration", "ground_truth": "{\"task_id\": \"mlqa_de_hi__36474374__s001__tb131072__v005\", \"budget\": 2048, \"token_budget\": 131072, \"ppl_base\": 17.219207942961326, \"nll_base\": 2.846025501598011, \"eval_ref\": \"mlqa_de_hi__36474374__s001__tb131072__v005\", \"oracle_delta\": 0.43866155450994304...
{ "index": "mlqa_de_hi__36474374__s001__tb131072__v005", "uid": "mlqa_de_hi__36474374__s001__tb131072__v005", "axis": "qa", "budget": 2048, "token_budget": 131072, "variant": "filter" }