Datasets:

ctjones
/

wordchain-rl-generations

query stringlengths 687 701	start_word stringlengths 3 10	end_word stringlengths 3 10	id stringlengths 7 10	messages listlengths 2 2	expert_reasoning stringlengths 1.89k 52.4k	expert_response stringlengths 246 3.06k	completion stringlengths 2.6k 53.4k	attempt int64 1 4
Make a word chain from "CLEAN" to "INVEST". The goal is to make the shortest valid chain. Each pair of adjacent words must be part of the same set phrase. Set phrases are phrases like "fast food," "beat around the bush," or "long time no see" whose meanings cannot be fully derived from their individual words. The word...	CLEAN	INVEST	train-0	[ { "role": "system", "content": "Your input fields are:\n1. `query` (str): The query\nYour output fields are:\n1. `response` (str): The response to the query\nAll interactions will be structured in the following way, with the appropriate values filled in.\n\n[[ ## query ## ]]\n{query}\n\n[[ ## response ## ]]...	Okay, I need to create the shortest word chain from "CLEAN" to "INVEST". The rule is that each adjacent pair must be part of a set phrase. Set phrases can have words in either order. So, first, I should figure out what set phrases include "CLEAN" and "INVEST", and then see how to connect them. Let me start by recallin...	ANSWER: CLEAN -> HOUSE -> IN -> INVEST Set phrases: - CLEAN and HOUSE: "clean house" - HOUSE and IN: "in the house" - IN and INVEST: "invest in" Critique: - CLEAN -> HOUSE: Valid. "Clean house" is a common set phrase meaning to tidy or purify a house. - HOUSE -> IN: Valid. "In the house" is a standard set phrase, and ...	<think> Okay, I need to create the shortest word chain from "CLEAN" to "INVEST". The rule is that each adjacent pair must be part of a set phrase. Set phrases can have words in either order. So, first, I should figure out what set phrases include "CLEAN" and "INVEST", and then see how to connect them. Let me start by ...	1
Make a word chain from "JUDGMENT" to "JOIN". The goal is to make the shortest valid chain. Each pair of adjacent words must be part of the same set phrase. Set phrases are phrases like "fast food," "beat around the bush," or "long time no see" whose meanings cannot be fully derived from their individual words. The wor...	JUDGMENT	JOIN	train-1	[ { "role": "system", "content": "Your input fields are:\n1. `query` (str): The query\nYour output fields are:\n1. `response` (str): The response to the query\nAll interactions will be structured in the following way, with the appropriate values filled in.\n\n[[ ## query ## ]]\n{query}\n\n[[ ## response ## ]]...	Okay, I need to create the shortest word chain from "JUDGMENT" to "JOIN". The rules are that each adjacent pair must be part of a set phrase, and the words in the set phrase can be in either order. Let me start by recalling what set phrases are. They are common expressions where the meaning isn't just the sum of the wo...	ANSWER: JUDGMENT -> CALL -> IN -> JOIN Set phrases: - JUDGMENT and CALL: "judgment call" - CALL and IN: "call in" - IN and JOIN: "join in" Critique: 1. JUDGMENT -> CALL: Valid. "Judgment call" is a recognized set phrase, referring to a decision based on judgment (e.g., "It was a judgment call"). 2. CALL -> IN:...	<think> Okay, I need to create the shortest word chain from "JUDGMENT" to "JOIN". The rules are that each adjacent pair must be part of a set phrase, and the words in the set phrase can be in either order. Let me start by recalling what set phrases are. They are common expressions where the meaning isn't just the sum o...	1
Make a word chain from "DADDY" to "PLOT". The goal is to make the shortest valid chain. Each pair of adjacent words must be part of the same set phrase. Set phrases are phrases like "fast food," "beat around the bush," or "long time no see" whose meanings cannot be fully derived from their individual words. The words ...	DADDY	PLOT	train-2	[ { "role": "system", "content": "Your input fields are:\n1. `query` (str): The query\nYour output fields are:\n1. `response` (str): The response to the query\nAll interactions will be structured in the following way, with the appropriate values filled in.\n\n[[ ## query ## ]]\n{query}\n\n[[ ## response ## ]]...	Okay, I need to create the shortest word chain from "DADDY" to "PLOT". The rules are that each adjacent pair must be part of a set phrase, and the words in the set phrase can be in either order. Let me start by recalling what set phrases are. They are common expressions where the words are used together, like "fast foo...	ANSWER: DADDY -> LAND -> PLOT Set phrases: - DADDY and LAND: "daddy's land" - LAND and PLOT: "plot of land" Critique: - "daddy's land": This phrase is not a commonly recognized set phrase. While "daddy's" can precede a noun (e.g., "daddy's car"), "land" is not typically associated with "daddy" in standard set phrases....	<think> Okay, I need to create the shortest word chain from "DADDY" to "PLOT". The rules are that each adjacent pair must be part of a set phrase, and the words in the set phrase can be in either order. Let me start by recalling what set phrases are. They are common expressions where the words are used together, like "...	1
Make a word chain from "SPANISH" to "HOPE". The goal is to make the shortest valid chain. Each pair of adjacent words must be part of the same set phrase. Set phrases are phrases like "fast food," "beat around the bush," or "long time no see" whose meanings cannot be fully derived from their individual words. The word...	SPANISH	HOPE	train-3	[ { "role": "system", "content": "Your input fields are:\n1. `query` (str): The query\nYour output fields are:\n1. `response` (str): The response to the query\nAll interactions will be structured in the following way, with the appropriate values filled in.\n\n[[ ## query ## ]]\n{query}\n\n[[ ## response ## ]]...	Okay, I need to create the shortest word chain from "SPANISH" to "HOPE". The rule is that each adjacent pair must be part of a set phrase, and the words can be in either order. Let me start by recalling what set phrases are. They are common expressions where the words are used together, like "fast food" or "beat around...	ANSWER: SPANISH -> FOR -> HOPE Set phrases: - SPANISH and FOR: "Spanish for [something]" (e.g., "Spanish for beginners") - FOR and HOPE: "hope for" (e.g., "hope for the best") Critique: - SPANISH -> FOR: Valid. "Spanish for" is a common set phrase (e.g., "Spanish for travelers"). - FOR -> HOPE: Valid. "Hope f...	<think> Okay, I need to create the shortest word chain from "SPANISH" to "HOPE". The rule is that each adjacent pair must be part of a set phrase, and the words can be in either order. Let me start by recalling what set phrases are. They are common expressions where the words are used together, like "fast food" or "bea...	1
"Make a word chain from \"SEA\" to \"LOL\".\nThe goal is to make the shortest valid chain.\n\nEach p(...TRUNCATED)	SEA	LOL	train-4	[{"role":"system","content":"Your input fields are:\n1. `query` (str): The query\nYour output fields(...TRUNCATED)	"Okay, I need to create the shortest word chain from \"SEA\" to \"LOL\". The rules are that each adj(...TRUNCATED)	"ANSWER: SEA -> at -> laugh -> LOL\nSet phrases:\n- SEA and at: \"at sea\"\n- at and laugh: \"laugh (...TRUNCATED)	"<think>\nOkay, I need to create the shortest word chain from \"SEA\" to \"LOL\". The rules are that(...TRUNCATED)	1
"Make a word chain from \"ANALYSIS\" to \"SOCIETY\".\nThe goal is to make the shortest valid chain.\(...TRUNCATED)	ANALYSIS	SOCIETY	train-5	[{"role":"system","content":"Your input fields are:\n1. `query` (str): The query\nYour output fields(...TRUNCATED)	"Okay, I need to create the shortest word chain from \"ANALYSIS\" to \"SOCIETY\". The rule is that e(...TRUNCATED)	"ANSWER: ANALYSIS -> SOCIAL -> ASPECTS -> SOCIETY\nSet phrases:\n- \"social analysis\" (ANALYSIS and(...TRUNCATED)	"<think>\nOkay, I need to create the shortest word chain from \"ANALYSIS\" to \"SOCIETY\". The rule (...TRUNCATED)	2
"Make a word chain from \"TALENT\" to \"PART\".\nThe goal is to make the shortest valid chain.\n\nEa(...TRUNCATED)	TALENT	PART	train-6	[{"role":"system","content":"Your input fields are:\n1. `query` (str): The query\nYour output fields(...TRUNCATED)	"Okay, I need to create a word chain from \"TALENT\" to \"PART\" using set phrases. The goal is the (...TRUNCATED)	"ANSWER: TALENT -> HAVE -> PART\nSet phrases:\n- TALENT and HAVE are in \"have a talent\" (or \"a ta(...TRUNCATED)	"<think>\nOkay, I need to create a word chain from \"TALENT\" to \"PART\" using set phrases. The goa(...TRUNCATED)	1
"Make a word chain from \"FOUNDER\" to \"THEN\".\nThe goal is to make the shortest valid chain.\n\nE(...TRUNCATED)	FOUNDER	THEN	train-7	[{"role":"system","content":"Your input fields are:\n1. `query` (str): The query\nYour output fields(...TRUNCATED)	"Okay, so I need to make a word chain from \"FOUNDER\" to \"THEN\". The goal is the shortest possibl(...TRUNCATED)	"ANSWER: FOUNDER -> AND -> THEN\nSet phrases:\n- FOUNDER and AND: from phrases like \"founder and CE(...TRUNCATED)	"<think>\nOkay, so I need to make a word chain from \"FOUNDER\" to \"THEN\". The goal is the shortes(...TRUNCATED)	1
"Make a word chain from \"SHED\" to \"BURDEN\".\nThe goal is to make the shortest valid chain.\n\nEa(...TRUNCATED)	SHED	BURDEN	train-8	[{"role":"system","content":"Your input fields are:\n1. `query` (str): The query\nYour output fields(...TRUNCATED)	"Okay, I need to create a word chain from \"SHED\" to \"BURDEN\" using set phrases. The goal is the (...TRUNCATED)	"ANSWER: SHED -> BURDEN\nSet phrases: \"shed the burden\"\n\nThe chain SHED → BURDEN is valid beca(...TRUNCATED)	"<think>\nOkay, I need to create a word chain from \"SHED\" to \"BURDEN\" using set phrases. The goa(...TRUNCATED)	1
"Make a word chain from \"WHAT\" to \"CHOOSE\".\nThe goal is to make the shortest valid chain.\n\nEa(...TRUNCATED)	WHAT	CHOOSE	train-9	[{"role":"system","content":"Your input fields are:\n1. `query` (str): The query\nYour output fields(...TRUNCATED)	"Okay, so I need to make a word chain from \"WHAT\" to \"CHOOSE\". The goal is the shortest possible(...TRUNCATED)	"ANSWER: WHAT -> CHOOSE\nSet phrases: \"choose what\" (connects WHAT and CHOOSE)\n\nCritique:\n- The(...TRUNCATED)	"<think>\nOkay, so I need to make a word chain from \"WHAT\" to \"CHOOSE\". The goal is the shortest(...TRUNCATED)	1

End of preview. Expand in Data Studio

Wordchain RL teacher generations

Reasoning and final-answer samples generated by Qwen/Qwen3-30B-A3B with the CDBiddulph/qwen3-30b-a3b-wordchain-rl adapter. These are generated distillation targets, rather than model weights or newly authored task inputs.

Split	Published rows	Source inputs	Excluded
train	9,999	10,000	1
validation	100	100	0
test	100	100	0

from datasets import load_dataset
data = load_dataset("ctjones/wordchain-rl-generations")

Download the Parquet files directly from Files and versions if preferred.

Fields

id: source split and row index; original validation IDs retain the valid- prefix.
query, start_word, end_word: original task input.
messages: exact system/user messages used for inference, without an assistant completion.
expert_reasoning, expert_response: parsed reasoning and final answer from the teacher. The field name expert identifies the generating model; it does not imply correctness.
completion: complete generated assistant text, including thinking and DSPy formatting.
attempt: selected generation attempt (1 through 4).

Generation and selection

Thinking enabled; temperature 1; initial seed 42; batches of 32; initial output limit 8,192 tokens; format/length retries limited to 16,000 output tokens. The final 27 rejected examples were allowed attempts 3 and 4, stopping after acceptance. Batched retries use per-example deterministic seeds; see the pinned code and provenance for the seed formulas.

Acceptance requires parsed, nonempty reasoning and answer plus a normal stop. There is no reward or task-correctness filtering. One training input remained rejected after four attempts and is absent from the training split; its ID and finish reason are in provenance/train_excluded.json. This repository contains the selected accepted generations, not every raw attempt.

Provenance

Base: Qwen/Qwen3-30B-A3B, revision ad44e777bcd18fa416d9da3bd8f70d33ebb85d39.
Adapter: CDBiddulph/qwen3-30b-a3b-wordchain-rl, revision c816b3b9f59bde7500af1833af8ae6e8c2e3a95d.
Inputs: CDBiddulph/wordchain-rl, revision de7fb2b4b39a9602c40a64b39f9678e97f32379c.
Code: verbalizing-subliminal-learning, fork_experiments/2026-10-21-000-wordchain-data/.
Initial generation commit: 1315e0301752092105bef31911f65397d8b25d00.
Batched retry/resume commit: 652d3e3116dcfb98c9be318497c2ff4fb54689d4.
Extra retry commit: e0d6299988861d9ba3e6ac9e8481f8eb3f395f5d.

provenance/ contains source manifests, export hashes, retry records, and SHA-256 checksums of the published Parquet files. Source JSONL hashes were verified before conversion, and every converted row was checked against its source.

License and limitations

Apache 2.0, following the upstream inputs and adapter. Credit the upstream dataset/model authors and retain their applicable notices when redistributing. Model-generated reasoning may be inaccurate and may not faithfully explain the model's behavior. Format acceptance is not a behavioral evaluation.

Downloads last month: 35