Datasets:

abhayesian
/

ryan-simulator-eval-v1-stylepref-judge-calls-tinker-base

Modalities:

Text

Formats:

Size:

Libraries:

Dataset card Data Studio Files Files and versions

xet

Community

Dataset Viewer

Auto-converted to Parquet Duplicate

Split (1)

train · 525 rows

judge stringclasses 3 values	mode stringclasses 1 value	prompt stringlengths 3.26k 10.4k	response stringlengths 0 1.26k	parsed stringlengths 49 1.12k	swapped bool 2 classes	substance_winner stringclasses 4 values	style_winner stringclasses 4 values	label_left stringclasses 1 value	label_right stringclasses 1 value	id stringlengths 63 83	pairing stringclasses 1 value	meta stringlengths 109 129
claude-opus	pairwise	You are a blind evaluator comparing two continuations to the same prompt. The prompt is the opening of a LessWrong post or comment by an unidentified AI-safety author. Your job is to decide which continuation is more likely the AUTHOR'S OWN continuation rather than a generic LLM completion. You will report TWO separat...	{ "substance_verdict": "A", "substance_rationale": "Continuation A provides concrete methodological context by referencing specific related work (DEI-RL, PPO) and situating the post within a sequence of prior work. Continuation B offers only generic descriptions of the approach without any technical depth or novel ...	{"substance_verdict": "A", "substance_rationale": "Continuation A provides concrete methodological context by referencing specific related work (DEI-RL, PPO) and situating the post within a sequence of prior work. Continuation B offers only generic descriptions of the approach without any technical depth or novel insig...	false	sft	sft	sft	tinker_base	sft__format_continuation__lr5e-05__step96__sp_000__sft_vs_tinker_base	sft_vs_tinker_base	{"checkpoint": "sft__format_continuation__lr5e-05__step96", "prompt_id": "sp_000", "pairing": "sft_vs_tinker_base"}
gpt-5	pairwise	You are a blind evaluator comparing two continuations to the same prompt. The prompt is the opening of a LessWrong post or comment by an unidentified AI-safety author. Your job is to decide which continuation is more likely the AUTHOR'S OWN continuation rather than a generic LLM completion. You will report TWO separat...	{ "substance_verdict": "B", "substance_rationale": "B outlines concrete mechanisms—example-based evaluation, heuristic filtering, iterative refinement—and sketches operational tweaks beyond brute-force sampling. A is mostly meta-context and references, with little detail on the actual method.", "style_verdict": "...	{"substance_verdict": "B", "substance_rationale": "B outlines concrete mechanisms\u2014example-based evaluation, heuristic filtering, iterative refinement\u2014and sketches operational tweaks beyond brute-force sampling. A is mostly meta-context and references, with little detail on the actual method.", "style_verdict"...	false	tinker_base	sft	sft	tinker_base	sft__format_continuation__lr5e-05__step96__sp_000__sft_vs_tinker_base	sft_vs_tinker_base	{"checkpoint": "sft__format_continuation__lr5e-05__step96", "prompt_id": "sp_000", "pairing": "sft_vs_tinker_base"}
gemini-3-pro	pairwise	You are a blind evaluator comparing two continuations to the same prompt. The prompt is the opening of a LessWrong post or comment by an unidentified AI-safety author. Your job is to decide which continuation is more likely the AUTHOR'S OWN continuation rather than a generic LLM completion. You will report TWO separat...	```json { "substance_verdict": "A", "substance_rationale": "Continuation A situates the work technically by contrasting it with RL methods and providing a roadmap for the post. Continuation B is filled with low-information, generic filler that explains trivial concepts (like why evaluating on examples is useful) wh...	{"substance_verdict": "A", "substance_rationale": "Continuation A situates the work technically by contrasting it with RL methods and providing a roadmap for the post. Continuation B is filled with low-information, generic filler that explains trivial concepts (like why evaluating on examples is useful) which would be ...	false	sft	sft	sft	tinker_base	sft__format_continuation__lr5e-05__step96__sp_000__sft_vs_tinker_base	sft_vs_tinker_base	{"checkpoint": "sft__format_continuation__lr5e-05__step96", "prompt_id": "sp_000", "pairing": "sft_vs_tinker_base"}
claude-opus	pairwise	You are a blind evaluator comparing two continuations to the same prompt. The prompt is the opening of a LessWrong post or comment by an unidentified AI-safety author. Your job is to decide which continuation is more likely the AUTHOR'S OWN continuation rather than a generic LLM completion. You will report TWO separat...	{ "substance_verdict": "tie", "substance_rationale": "Neither continuation provides strong substantive AI-safety reasoning. Continuation A focuses on meta-discussion about what counts as SoTA without explaining the actual method. Continuation B provides a clearer breakdown of the approach but is generic and doesn't...	{"substance_verdict": "tie", "substance_rationale": "Neither continuation provides strong substantive AI-safety reasoning. Continuation A focuses on meta-discussion about what counts as SoTA without explaining the actual method. Continuation B provides a clearer breakdown of the approach but is generic and doesn't enga...	false	tie	sft	sft	tinker_base	sft__format_dialogue__lr5e-05__step140__sp_000__sft_vs_tinker_base	sft_vs_tinker_base	{"checkpoint": "sft__format_dialogue__lr5e-05__step140", "prompt_id": "sp_000", "pairing": "sft_vs_tinker_base"}
gpt-5	pairwise	You are a blind evaluator comparing two continuations to the same prompt. The prompt is the opening of a LessWrong post or comment by an unidentified AI-safety author. Your job is to decide which continuation is more likely the AUTHOR'S OWN continuation rather than a generic LLM completion. You will report TWO separat...	{ "substance_verdict": "A", "substance_rationale": "A engages with evaluation methodology and SOTA criteria, discusses data leakage/overlap (10k public examples; 5k in test), and clarifies the bar between train vs test—concrete issues that matter for credible claims. B is a generic methodological summary with no gr...	{"substance_verdict": "A", "substance_rationale": "A engages with evaluation methodology and SOTA criteria, discusses data leakage/overlap (10k public examples; 5k in test), and clarifies the bar between train vs test\u2014concrete issues that matter for credible claims. B is a generic methodological summary with no gr...	false	sft	sft	sft	tinker_base	sft__format_dialogue__lr5e-05__step140__sp_000__sft_vs_tinker_base	sft_vs_tinker_base	{"checkpoint": "sft__format_dialogue__lr5e-05__step140", "prompt_id": "sp_000", "pairing": "sft_vs_tinker_base"}
gemini-3-pro	pairwise	You are a blind evaluator comparing two continuations to the same prompt. The prompt is the opening of a LessWrong post or comment by an unidentified AI-safety author. Your job is to decide which continuation is more likely the AUTHOR'S OWN continuation rather than a generic LLM completion. You will report TWO separat...	```json { "substance_verdict": "B", "substance_rationale": "Neither continuation provides deep technical insights, but Continuation B offers a coherent, albeit generic, breakdown of plausible mechanisms for the 'additional approaches and tweaks' (e.g., heuristic filtering, iterative improvement). Continuation A, by...	{"substance_verdict": "B", "substance_rationale": "Neither continuation provides deep technical insights, but Continuation B offers a coherent, albeit generic, breakdown of plausible mechanisms for the 'additional approaches and tweaks' (e.g., heuristic filtering, iterative improvement). Continuation A, by contrast, is...	false	tinker_base	sft	sft	tinker_base	sft__format_dialogue__lr5e-05__step140__sp_000__sft_vs_tinker_base	sft_vs_tinker_base	{"checkpoint": "sft__format_dialogue__lr5e-05__step140", "prompt_id": "sp_000", "pairing": "sft_vs_tinker_base"}
claude-opus	pairwise	You are a blind evaluator comparing two continuations to the same prompt. The prompt is the opening of a LessWrong post or comment by an unidentified AI-safety author. Your job is to decide which continuation is more likely the AUTHOR'S OWN continuation rather than a generic LLM completion. You will report TWO separat...	{ "substance_verdict": "tie", "substance_rationale": "Both continuations describe the same basic approach (generate many solutions, filter by example correctness) without providing concrete technical details about the 'additional approaches and tweaks' mentioned in the prompt. Neither engages with specific threat m...	{"substance_verdict": "tie", "substance_rationale": "Both continuations describe the same basic approach (generate many solutions, filter by example correctness) without providing concrete technical details about the 'additional approaches and tweaks' mentioned in the prompt. Neither engages with specific threat models...	false	tie	sft	sft	tinker_base	sft__format_distillation__lr5e-05__step80__sp_000__sft_vs_tinker_base	sft_vs_tinker_base	{"checkpoint": "sft__format_distillation__lr5e-05__step80", "prompt_id": "sp_000", "pairing": "sft_vs_tinker_base"}
gpt-5	pairwise	You are a blind evaluator comparing two continuations to the same prompt. The prompt is the opening of a LessWrong post or comment by an unidentified AI-safety author. Your job is to decide which continuation is more likely the AUTHOR'S OWN continuation rather than a generic LLM completion. You will report TWO separat...	{ "substance_verdict": "B", "substance_rationale": "B offers a clearer algorithmic breakdown, including heuristic filtering and iterative improvement using best-performing variants, which is more substantive than A’s generic points. While still high-level, it at least sketches mechanisms beyond mere sampling and se...	{"substance_verdict": "B", "substance_rationale": "B offers a clearer algorithmic breakdown, including heuristic filtering and iterative improvement using best-performing variants, which is more substantive than A\u2019s generic points. While still high-level, it at least sketches mechanisms beyond mere sampling and se...	false	tinker_base	sft	sft	tinker_base	sft__format_distillation__lr5e-05__step80__sp_000__sft_vs_tinker_base	sft_vs_tinker_base	{"checkpoint": "sft__format_distillation__lr5e-05__step80", "prompt_id": "sp_000", "pairing": "sft_vs_tinker_base"}
gemini-3-pro	pairwise	You are a blind evaluator comparing two continuations to the same prompt. The prompt is the opening of a LessWrong post or comment by an unidentified AI-safety author. Your job is to decide which continuation is more likely the AUTHOR'S OWN continuation rather than a generic LLM completion. You will report TWO separat...	```json { "substance_verdict": "tie", "substance_rationale": "Both continuations are generic LLM completions that fail to provide any concrete, substantive details about the 'additional approaches and tweaks' mentioned in the prompt. Instead of engaging with the actual mechanics of the ARC-AGI solution, both rely o...	{"substance_verdict": "tie", "substance_rationale": "Both continuations are generic LLM completions that fail to provide any concrete, substantive details about the 'additional approaches and tweaks' mentioned in the prompt. Instead of engaging with the actual mechanics of the ARC-AGI solution, both rely on hallucinate...	false	tie	tie	sft	tinker_base	sft__format_distillation__lr5e-05__step80__sp_000__sft_vs_tinker_base	sft_vs_tinker_base	{"checkpoint": "sft__format_distillation__lr5e-05__step80", "prompt_id": "sp_000", "pairing": "sft_vs_tinker_base"}
claude-opus	pairwise	You are a blind evaluator comparing two continuations to the same prompt. The prompt is the opening of a LessWrong post or comment by an unidentified AI-safety author. Your job is to decide which continuation is more likely the AUTHOR'S OWN continuation rather than a generic LLM completion. You will report TWO separat...	{ "substance_verdict": "tie", "substance_rationale": "Neither continuation provides strong substantive AI-safety reasoning. Continuation A contains confusing and contradictory statements about human performance statistics that don't make sense. Continuation B provides a generic overview of the method without any no...	{"substance_verdict": "tie", "substance_rationale": "Neither continuation provides strong substantive AI-safety reasoning. Continuation A contains confusing and contradictory statements about human performance statistics that don't make sense. Continuation B provides a generic overview of the method without any novel i...	false	tie	sft	sft	tinker_base	sft__format_qa_v2__lr5e-05__step100__sp_000__sft_vs_tinker_base	sft_vs_tinker_base	{"checkpoint": "sft__format_qa_v2__lr5e-05__step100", "prompt_id": "sp_000", "pairing": "sft_vs_tinker_base"}
gpt-5	pairwise	You are a blind evaluator comparing two continuations to the same prompt. The prompt is the opening of a LessWrong post or comment by an unidentified AI-safety author. Your job is to decide which continuation is more likely the AUTHOR'S OWN continuation rather than a generic LLM completion. You will report TWO separat...	{ "substance_verdict": "B", "substance_rationale": "B gives a coherent breakdown of the method (generation, example-based selection, heuristics, iterative refinement) and plausible extensions, whereas A is self-contradictory about human performance and adds no concrete mechanism or evaluation detail. B, though gene...	{"substance_verdict": "B", "substance_rationale": "B gives a coherent breakdown of the method (generation, example-based selection, heuristics, iterative refinement) and plausible extensions, whereas A is self-contradictory about human performance and adds no concrete mechanism or evaluation detail. B, though generic, ...	false	tinker_base	tinker_base	sft	tinker_base	sft__format_qa_v2__lr5e-05__step100__sp_000__sft_vs_tinker_base	sft_vs_tinker_base	{"checkpoint": "sft__format_qa_v2__lr5e-05__step100", "prompt_id": "sp_000", "pairing": "sft_vs_tinker_base"}
gemini-3-pro	pairwise	You are a blind evaluator comparing two continuations to the same prompt. The prompt is the opening of a LessWrong post or comment by an unidentified AI-safety author. Your job is to decide which continuation is more likely the AUTHOR'S OWN continuation rather than a generic LLM completion. You will report TWO separat...	```json { "substance_verdict": "tie", "substance_rationale": "Neither continuation provides meaningful technical or AI-safety reasoning. Continuation A is mathematically incoherent (confusing median and 50th percentile), while Continuation B is pure LLM filler that merely restates the prompt using generic analogies...	{"substance_verdict": "tie", "substance_rationale": "Neither continuation provides meaningful technical or AI-safety reasoning. Continuation A is mathematically incoherent (confusing median and 50th percentile), while Continuation B is pure LLM filler that merely restates the prompt using generic analogies and lists wi...	false	tie	sft	sft	tinker_base	sft__format_qa_v2__lr5e-05__step100__sp_000__sft_vs_tinker_base	sft_vs_tinker_base	{"checkpoint": "sft__format_qa_v2__lr5e-05__step100", "prompt_id": "sp_000", "pairing": "sft_vs_tinker_base"}
claude-opus	pairwise	You are a blind evaluator comparing two continuations to the same prompt. The prompt is the opening of a LessWrong post or comment by an unidentified AI-safety author. Your job is to decide which continuation is more likely the AUTHOR'S OWN continuation rather than a generic LLM completion. You will report TWO separat...	{ "substance_verdict": "A", "substance_rationale": "Continuation A promises concrete content (ablation study, method selection rationale) and acknowledges limitations via a linked comment. Continuation B is generic and explanatory, restating the prompt's approach with vague 'future directions' rather than presentin...	{"substance_verdict": "A", "substance_rationale": "Continuation A promises concrete content (ablation study, method selection rationale) and acknowledges limitations via a linked comment. Continuation B is generic and explanatory, restating the prompt's approach with vague 'future directions' rather than presenting act...	false	sft	sft	sft	tinker_base	sft__mix_comment_deduped__lr2e-04__step300__sp_000__sft_vs_tinker_base	sft_vs_tinker_base	{"checkpoint": "sft__mix_comment_deduped__lr2e-04__step300", "prompt_id": "sp_000", "pairing": "sft_vs_tinker_base"}
gpt-5	pairwise	You are a blind evaluator comparing two continuations to the same prompt. The prompt is the opening of a LessWrong post or comment by an unidentified AI-safety author. Your job is to decide which continuation is more likely the AUTHOR'S OWN continuation rather than a generic LLM completion. You will report TWO separat...	{ "substance_verdict": "A", "substance_rationale": "A gestures at concrete evaluations (an ablation study), method selection tradeoffs, and explicitly flags limits on what the result says about GPT-4o’s capabilities. B is generic process description without identifying cruxes, limits, or operational tests specific ...	{"substance_verdict": "A", "substance_rationale": "A gestures at concrete evaluations (an ablation study), method selection tradeoffs, and explicitly flags limits on what the result says about GPT-4o\u2019s capabilities. B is generic process description without identifying cruxes, limits, or operational tests specific ...	false	sft	sft	sft	tinker_base	sft__mix_comment_deduped__lr2e-04__step300__sp_000__sft_vs_tinker_base	sft_vs_tinker_base	{"checkpoint": "sft__mix_comment_deduped__lr2e-04__step300", "prompt_id": "sp_000", "pairing": "sft_vs_tinker_base"}
gemini-3-pro	pairwise	You are a blind evaluator comparing two continuations to the same prompt. The prompt is the opening of a LessWrong post or comment by an unidentified AI-safety author. Your job is to decide which continuation is more likely the AUTHOR'S OWN continuation rather than a generic LLM completion. You will report TWO separat...	```json { "substance_verdict": "A", "substance_rationale": "Continuation A sets up a concrete roadmap for the post, including an ablation study and a discussion on the limits of what the experiment implies about model capabilities. Continuation B merely regurgitates the prompt using generic machine learning filler ...	{"substance_verdict": "A", "substance_rationale": "Continuation A sets up a concrete roadmap for the post, including an ablation study and a discussion on the limits of what the experiment implies about model capabilities. Continuation B merely regurgitates the prompt using generic machine learning filler (e.g., 'Fine-...	false	sft	sft	sft	tinker_base	sft__mix_comment_deduped__lr2e-04__step300__sp_000__sft_vs_tinker_base	sft_vs_tinker_base	{"checkpoint": "sft__mix_comment_deduped__lr2e-04__step300", "prompt_id": "sp_000", "pairing": "sft_vs_tinker_base"}
claude-opus	pairwise	You are a blind evaluator comparing two continuations to the same prompt. The prompt is the opening of a LessWrong post or comment by an unidentified AI-safety author. Your job is to decide which continuation is more likely the AUTHOR'S OWN continuation rather than a generic LLM completion. You will report TWO separat...	{ "substance_verdict": "A", "substance_rationale": "Continuation A provides more concrete technical details about the experimental setup (single run, no CoT, no finetuning), links to code/resources, and outlines a specific structure for the rest of the post. While it contains an incorrect description of ARC-AGI at ...	{"substance_verdict": "A", "substance_rationale": "Continuation A provides more concrete technical details about the experimental setup (single run, no CoT, no finetuning), links to code/resources, and outlines a specific structure for the rest of the post. While it contains an incorrect description of ARC-AGI at the e...	false	sft	sft	sft	tinker_base	sft__mix_examples_balanced_tokmatched__lr2e-04__step200__sp_000__sft_vs_tinker_base	sft_vs_tinker_base	{"checkpoint": "sft__mix_examples_balanced_tokmatched__lr2e-04__step200", "prompt_id": "sp_000", "pairing": "sft_vs_tinker_base"}
gpt-5	pairwise	You are a blind evaluator comparing two continuations to the same prompt. The prompt is the opening of a LessWrong post or comment by an unidentified AI-safety author. Your job is to decide which continuation is more likely the AUTHOR'S OWN continuation rather than a generic LLM completion. You will report TWO separat...	{ "substance_verdict": "B", "substance_rationale": "A contains clear inaccuracies and implausible claims (e.g., incorrect description of ARC-AGI and dubious compute/link details), while B gives a coherent, on-topic breakdown of the generate-and-select method with plausible refinements. Even if generic, B’s content ...	{"substance_verdict": "B", "substance_rationale": "A contains clear inaccuracies and implausible claims (e.g., incorrect description of ARC-AGI and dubious compute/link details), while B gives a coherent, on-topic breakdown of the generate-and-select method with plausible refinements. Even if generic, B\u2019s content ...	false	tinker_base	sft	sft	tinker_base	sft__mix_examples_balanced_tokmatched__lr2e-04__step200__sp_000__sft_vs_tinker_base	sft_vs_tinker_base	{"checkpoint": "sft__mix_examples_balanced_tokmatched__lr2e-04__step200", "prompt_id": "sp_000", "pairing": "sft_vs_tinker_base"}
gemini-3-pro	pairwise	You are a blind evaluator comparing two continuations to the same prompt. The prompt is the opening of a LessWrong post or comment by an unidentified AI-safety author. Your job is to decide which continuation is more likely the AUTHOR'S OWN continuation rather than a generic LLM completion. You will report TWO separat...	```json { "substance_verdict": "tie", "substance_rationale": "Both continuations fail to provide meaningful technical substance. Continuation A attempts to provide concrete experimental details but hallucinates wildly (e.g., a completely incorrect definition of ARC-AGI and non-existent hardware), while Continuation...	{"substance_verdict": "tie", "substance_rationale": "Both continuations fail to provide meaningful technical substance. Continuation A attempts to provide concrete experimental details but hallucinates wildly (e.g., a completely incorrect definition of ARC-AGI and non-existent hardware), while Continuation B provides v...	false	tie	sft	sft	tinker_base	sft__mix_examples_balanced_tokmatched__lr2e-04__step200__sp_000__sft_vs_tinker_base	sft_vs_tinker_base	{"checkpoint": "sft__mix_examples_balanced_tokmatched__lr2e-04__step200", "prompt_id": "sp_000", "pairing": "sft_vs_tinker_base"}
claude-opus	pairwise	You are a blind evaluator comparing two continuations to the same prompt. The prompt is the opening of a LessWrong post or comment by an unidentified AI-safety author. Your job is to decide which continuation is more likely the AUTHOR'S OWN continuation rather than a generic LLM completion. You will report TWO separat...	{ "substance_verdict": "A", "substance_rationale": "Continuation A makes a substantive distinction between public test set performance and actual ARC-AGI solving capability, noting issues with sample efficiency and generalization. While both continuations are relatively shallow, A at least identifies a concrete lim...	{"substance_verdict": "A", "substance_rationale": "Continuation A makes a substantive distinction between public test set performance and actual ARC-AGI solving capability, noting issues with sample efficiency and generalization. While both continuations are relatively shallow, A at least identifies a concrete limitati...	false	sft	sft	sft	tinker_base	sft__format_continuation__lr2e-04__step96__sp_000__sft_vs_tinker_base	sft_vs_tinker_base	{"checkpoint": "sft__format_continuation__lr2e-04__step96", "prompt_id": "sp_000", "pairing": "sft_vs_tinker_base"}
gpt-5	pairwise	You are a blind evaluator comparing two continuations to the same prompt. The prompt is the opening of a LessWrong post or comment by an unidentified AI-safety author. Your job is to decide which continuation is more likely the AUTHOR'S OWN continuation rather than a generic LLM completion. You will report TWO separat...	{ "substance_verdict": "B", "substance_rationale": "B outlines the sampling-and-selection pipeline with concrete tweaks (prompting, heuristic filtering, iterative refinement) without obvious errors. A includes factual inaccuracies about ARC (e.g., only black/white, fixed 19x19 grids) and makes a dubious claim about...	{"substance_verdict": "B", "substance_rationale": "B outlines the sampling-and-selection pipeline with concrete tweaks (prompting, heuristic filtering, iterative refinement) without obvious errors. A includes factual inaccuracies about ARC (e.g., only black/white, fixed 19x19 grids) and makes a dubious claim about publ...	false	tinker_base	sft	sft	tinker_base	sft__format_continuation__lr2e-04__step96__sp_000__sft_vs_tinker_base	sft_vs_tinker_base	{"checkpoint": "sft__format_continuation__lr2e-04__step96", "prompt_id": "sp_000", "pairing": "sft_vs_tinker_base"}
gemini-3-pro	pairwise	You are a blind evaluator comparing two continuations to the same prompt. The prompt is the opening of a LessWrong post or comment by an unidentified AI-safety author. Your job is to decide which continuation is more likely the AUTHOR'S OWN continuation rather than a generic LLM completion. You will report TWO separat...	```json { "substance_verdict": "A", "substance_rationale": "Continuation A engages with specific, concrete concepts relevant to the ARC-AGI benchmark, such as the distinction between the public and private test sets and the sample efficiency of the approach. Continuation B relies on generic, high-level filler (e.g....	{"substance_verdict": "A", "substance_rationale": "Continuation A engages with specific, concrete concepts relevant to the ARC-AGI benchmark, such as the distinction between the public and private test sets and the sample efficiency of the approach. Continuation B relies on generic, high-level filler (e.g., 'akin to ha...	false	sft	sft	sft	tinker_base	sft__format_continuation__lr2e-04__step96__sp_000__sft_vs_tinker_base	sft_vs_tinker_base	{"checkpoint": "sft__format_continuation__lr2e-04__step96", "prompt_id": "sp_000", "pairing": "sft_vs_tinker_base"}

End of preview. Expand in Data Studio

README.md exists but content is empty.

Downloads last month: 12