Emulated-Inc/instruction-following-training-pool

Instruction following training pool

Public prompts for writing tasks, many of them carrying a constraint a program can check, from eight datasets read at the pinned revisions named below and one layer built here from them. The pool is laid out twice. Train on either layer or on both.

pool.jsonl

Every source rewritten into one shape, 310602 rows, one JSON object per line, with these fields.

Field	What it holds
id	a row identifier unique within this file
prompt	the request, as a model should receive it
response	an answer to it where the source ships one, null otherwise
instruction_id_list	the constraints the prompt carries, each named by the checker that decides it, empty when none were recovered
kwargs	one object per entry of that list, holding that checker's arguments
instruction_origin	how the constraints came to be attached: source, mined or built
instructions_complete	true when the listed constraints are every constraint the prompt states
source	the name of the source directory the row came from
source_repo, source_revision	the dataset and the revision it was read at
source_split, source_id	where the row sits in that set
provenance_class	how the row came to exist
licence	the licence of the source it came from

Of those rows 285170 carry a response and 27761 carry at least one constraint, 38821 constraints in all. A row with constraints can be scored without a reference answer, by running the named checker over a generated response with the arguments beside it, which is what the instruction_id_list and kwargs fields are for. A row with a response and no constraints is ordinary supervised data. Rows that carry both are the ones where a response was available to prove the constraint holds.

The three origins differ in what they guarantee. source means the set published the constraint itself, so it is exactly what the prompt asks. mined means the prompt states the constraint in words that a narrow set of patterns recognised and the row's own response already satisfies the checker, so the constraint is true of the response by construction, but the prompt may ask for other things besides. built means the request is a human one and the constraints after it were sampled and written at build time, so they are complete and exact but the wording is not human. instructions_complete records that difference per row, and the seed for the sampling was 20260911.

The constraints

Every constraint in the file is named by one of these checkers, with the arguments the name implies. A checker decides its constraint by reading the response, so the same vocabulary can be used as a training reward.

Checker	Rows
change_case:capital_word_frequency	1270
change_case:english_capital	1879
change_case:english_lowercase	1279
combination:repeat_prompt	896
combination:two_responses	986
detectable_content:number_placeholders	2070
detectable_content:postscript	1516
detectable_format:json_format	977
detectable_format:multiple_sections	1278
detectable_format:number_bullet_lists	1485
detectable_format:number_highlighted_sections	1280
detectable_format:title	1599
keywords:existence	2515
keywords:forbidden_words	2211
keywords:frequency	1279
keywords:letter_frequency	1412
language:response_language	1188
length_constraints:nth_paragraph_first_word	1303
length_constraints:number_paragraphs	1246
length_constraints:number_sentences	2746
length_constraints:number_words	3563
punctuation:no_comma	2072
startend:end_checker	1488
startend:quotation	1283

sources/

The same data untouched, 546342 rows, one directory per source, holding the files at the paths, in the formats and with the columns its own repository publishes. Nothing here was renamed, reshaped, reordered or deduplicated. Use this layer if you want a field the rewritten one drops, such as the later turns of a conversation, the reviewer rankings of the volunteer trees or the constraint descriptions of the reinforcement set, or if you would rather pair prompts with constraints yourself.

The sources

sources/dolly

Instruction and response pairs written by several thousand employees of one company across seven task categories, with reference passages copied from Wikipedia for the categories that need them. From databricks/databricks-dolly-15k at revision bdd27f4d94b9c1f951818a7da7fd7aeea5dbff1a, files databricks-dolly-15k.jsonl. 14993 rows here, of which 14776 also appear in pool.jsonl, 2 of them with constraints. Provenance class human, licence cc-by-sa-3.0. Its own fields are instruction, context, response, category.

Worth knowing. The contributors were told not to use a language model for either side, so both sides are human, but the writing is question answering and summarising rather than constrained writing, so almost no row states a constraint a checker can decide.

sources/oasst1

Message trees from a volunteer effort to build an assistant corpus by hand, in thirty-five languages, each message reviewed and ranked by other volunteers. From OpenAssistant/oasst1 at revision fdf72ae0827c1cda404aff25b6603abec9e3399b, files data/train-00000-of-00001-b42a775f407cee45.parquet, data/validation-00000-of-00001-134b8fd0c89408b6.parquet. 88810 rows here, of which 10226 also appear in pool.jsonl, 4 of them with constraints. Provenance class human, licence apache-2.0. Its own fields are message_id, parent_id, user_id, created_date, text, role, lang, review_count, review_result, deleted, rank, synthetic, model_name, detoxify, message_tree_id, tree_state, emojis, labels.

Worth knowing. The schema carries a synthetic flag and a model name, so a subset is model-written and is dropped from the rewritten rows. The rewritten rows take the first message of a tree and the reply the volunteers ranked first, so later turns live in this directory only.

sources/oasst2

The later release of the same volunteer message trees, with more languages. From OpenAssistant/oasst2 at revision 179dd21fc55192153d94adb0e0ce8f69e222bf75, files data/train-00000-of-00001-88ba0162028a73fc.parquet, data/validation-00000-of-00001-1deeef95c3248fe0.parquet. 135121 rows here, of which 3431 also appear in pool.jsonl, 1 of them with constraints. Provenance class human, licence apache-2.0. Its own fields are the same columns as the earlier release.

Worth knowing. It contains the earlier release, so the rewritten rows of the two overlap heavily and the duplicates are dropped in favour of the earlier one.

sources/rlvr_ifeval

Prompts drawn from an earlier supervised mixture with one machine-checkable constraint appended to each, shipped with the constraint's checker name and arguments and with no response. From allenai/RLVR-IFeval at revision 47c03c73621c4aab2b824b7818681117d662770e, files data/train-00000-of-00001.parquet. 5366 rows here, of which 5365 also appear in pool.jsonl, 2857 of them with constraints. Provenance class rule-generated, licence odc-by. Its own fields are messages, ground_truth, dataset, constraint_type, constraint.

Worth knowing. Its constraints are drawn from the same public registry that most of the sets here draw on, so its constraint sentences collide with the held-out prompts and two thirds of its rows were removed by the overlap rule. Four of its constraint kinds cannot be stated exactly in this vocabulary, and those rows keep the prompt with no constraints beside it.

sources/conifer

Conversations built by reframing collected user queries into instructions with layered constraints and answering them with a large proprietary model. From ConiferLM/Conifer at revision 8f085348895f8b794bbc8e1ccae705473625e156, files data/train_sft-00000-of-00001.parquet. 13606 rows here, of which 13576 also appear in pool.jsonl, 89 of them with constraints. Provenance class model-generated, licence apache-2.0. Its own fields are prompt, messages, type.

Worth knowing. Both sides are model-written, and only the seed queries trace to people. The constraints are stated in free prose rather than drawn from a registry, so what is recovered here is only what the mining patterns recognise.

sources/autoif

Conversations whose instructions were generated from hand-written seed instructions and collected queries, kept only when a generated checker function and its unit tests accepted the generated response. From Post-training-Data-Flywheel/AutoIF-instruct-61k at revision 0705aa867cce36d595123cb34f4a35281e6acc7c, files data/train-00000-of-00001.parquet. 59819 rows here, of which 37386 also appear in pool.jsonl, 600 of them with constraints. Provenance class model-generated, licence apache-2.0. Its own fields are system, tools, messages, conversation_id.

Worth knowing. The generated checkers are not shipped, so the constraints here are recovered by mining like any other prose set. The publishing account is not the laboratory that wrote the method.

sources/ultraif

Instructions composed by a constraint-composing model over collected seed queries, answered by an open 8B instruct model. From kkk-an/UltraIF-sft-175k at revision 1c14ef537756de2a41ee75a61c34711719f18d74, files UltraIF_llama31_8B_generated_sft_175k.jsonl. 181082 rows here, of which 178315 also appear in pool.jsonl, 875 of them with constraints. Provenance class model-generated, licence cc-by-4.0. Its own fields are instruction, output, history.

Worth knowing. Both sides are model-written. Rows carry a prior conversation in history, which the rewritten rows drop in favour of the final instruction and its answer, and a noticeable minority of the seed queries are not in English.

sources/personas_if

Prompts written by expanding a persona description into a request with two to four constraints attached, answered by a large model. From allenai/tulu-3-sft-personas-instruction-following at revision fe0c7d350c9b4542b8d829a6f1daa1c259f0ba0e, files data/train-00000-of-00001.parquet. 27479 rows here, of which 27461 also appear in pool.jsonl, 3267 of them with constraints. Provenance class model-generated, licence odc-by. Its own fields are id, prompt, messages, constraints.

Worth knowing. The constraint taxonomy is deliberately copied from the public registry, and the constraints ship as free-text lines rather than as checker arguments, so they are recovered here only by mining.

sources/built

A prompt of one of the human sets with one to three checkable instructions sampled from the registry and stated in words written for this pool, shipped with the instruction ids and the checker arguments and no response. Built here from the rows of the human sources above. 20066 rows here, of which 20066 also appear in pool.jsonl, 20066 of them with constraints. Provenance class rule-generated, licence inherited. Its own fields are prompt, seed_prompt, instruction_sentences, instruction_id_list, kwargs, source, source_repo, source_revision, source_id, licence.

Worth knowing. The request is human and the constraint sentences are not: they were written here and sampled at build time, which is why the class is rule-generated. Every row carries the licence of the set its request came from. Combinations the registry calls contradictory are never sampled together, but no response was written to prove an item answerable.

Provenance and licences

Every row carries the class of the way it came to exist, and the rewritten layer holds 28433 human, 256738 model-generated, 25431 rule-generated. Human means a person wrote the text. Model- generated means a language model wrote the prompt, the response or both, which is true of most published instruction-following data. Rule-generated means a program attached constraints to text it did not write. Nothing has been filtered out on its class: the field is there so that you can.

The pool as a whole is offered under cc-by-sa-4.0, which is the most restrictive term its sources compose to, since one human set is share-alike. Each rewritten row carries its own licence in the licence field and each directory under sources/ is one source, so a subset under a single licence can be selected. Attribution goes to the volunteers of the open assistant effort, to the employees who wrote the human instruction set, and to the authors of the generated and reinforcement sets, each named by its repository above.

Filtering

Rows whose prompt repeated a prompt of a held-out set, or shared an eight-word run with one, were removed before publication, from both layers alike: 0 rows on the first rule and 14774 on the second. That held-out set is not distributed here. The second rule is blunt on data of this kind, because a constraint stated in a standard sentence is an eight-word run in its own right, so the reinforcement set, whose constraints come from the same public registry, lost the greater part of its rows and the sets of ordinary human requests lost almost none. Nothing else was filtered: no subject, no language and no difficulty was selected for or against.