Datasets:

abhayesian
/

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

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 · 1.05k rows

judge stringclasses 3 values	mode stringclasses 1 value	prompt stringlengths 3.83k 11.3k	response stringlengths 0 1.26k	parsed stringlengths 49 1.62k	swapped bool 2 classes	substance_winner stringclasses 5 values	style_winner stringclasses 5 values	label_left stringclasses 1 value	label_right stringclasses 2 values	id stringlengths 56 80	pairing stringclasses 2 values	meta stringlengths 102 126
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": "B", "substance_rationale": "Continuation B provides concrete technical details about the methodology (multi-stage filtering, meta-evaluation, symbolic reasoning post-processing, problem-specific heuristics), discusses why these approaches work, and acknowledges limitations. Continuation A is...	{"substance_verdict": "B", "substance_rationale": "Continuation B provides concrete technical details about the methodology (multi-stage filtering, meta-evaluation, symbolic reasoning post-processing, problem-specific heuristics), discusses why these approaches work, and acknowledges limitations. Continuation A is most...	false	base	sft	sft	base	sft__format_continuation__lr5e-05__step96__sp_000__sft_vs_base	sft_vs_base	{"checkpoint": "sft__format_continuation__lr5e-05__step96", "prompt_id": "sp_000", "pairing": "sft_vs_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 concrete mechanisms (multi-stage filtering with example tests, LLM meta-ranking for generalizability, symbolic op library mapping, and a small guiding model) and grapples with overfitting and distributional caveats. A stays meta, invokes RL analogies that ...	{"substance_verdict": "B", "substance_rationale": "B gives concrete mechanisms (multi-stage filtering with example tests, LLM meta-ranking for generalizability, symbolic op library mapping, and a small guiding model) and grapples with overfitting and distributional caveats. A stays meta, invokes RL analogies that aren\...	false	base	base	sft	base	sft__format_continuation__lr5e-05__step96__sp_000__sft_vs_base	sft_vs_base	{"checkpoint": "sft__format_continuation__lr5e-05__step96", "prompt_id": "sp_000", "pairing": "sft_vs_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...	{ "substance_verdict": "A", "substance_rationale": "Continuation B proposes generic methods that mostly just repeat the prompt (filtering by example correctness) and explicitly contradicts the prompt's core premise by stating the author is 'far from the 50% SoTA threshold.' Continuation A provides a coherent struct...	{"substance_verdict": "A", "substance_rationale": "Continuation B proposes generic methods that mostly just repeat the prompt (filtering by example correctness) and explicitly contradicts the prompt's core premise by stating the author is 'far from the 50% SoTA threshold.' Continuation A provides a coherent structural ...	false	sft	sft	sft	base	sft__format_continuation__lr5e-05__step96__sp_000__sft_vs_base	sft_vs_base	{"checkpoint": "sft__format_continuation__lr5e-05__step96", "prompt_id": "sp_000", "pairing": "sft_vs_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 engages with concrete mechanisms (syntactic validity rates, overfitting risks, diversity penalties) and identifies specific failure modes (evaluation gaming, memorization vs. generalization). It proposes testable claims with rough percentages and ac...	{"substance_verdict": "A", "substance_rationale": "Continuation A engages with concrete mechanisms (syntactic validity rates, overfitting risks, diversity penalties) and identifies specific failure modes (evaluation gaming, memorization vs. generalization). It proposes testable claims with rough percentages and acknowl...	true	prompted_base	sft	sft	prompted_base	sft__format_continuation__lr5e-05__step96__sp_000__sft_vs_prompted	sft_vs_prompted	{"checkpoint": "sft__format_continuation__lr5e-05__step96", "prompt_id": "sp_000", "pairing": "sft_vs_prompted"}
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 gives concrete empirical estimates, discusses overfitting/cheating mechanisms, and describes specific heuristics (token reuse/diversity penalty) tied to generalization concerns. B is mostly logistics and prior-links with little engagement on failure modes or loa...	{"substance_verdict": "A", "substance_rationale": "A gives concrete empirical estimates, discusses overfitting/cheating mechanisms, and describes specific heuristics (token reuse/diversity penalty) tied to generalization concerns. B is mostly logistics and prior-links with little engagement on failure modes or load-bea...	true	prompted_base	prompted_base	sft	prompted_base	sft__format_continuation__lr5e-05__step96__sp_000__sft_vs_prompted	sft_vs_prompted	{"checkpoint": "sft__format_continuation__lr5e-05__step96", "prompt_id": "sp_000", "pairing": "sft_vs_prompted"}
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 attempts to engage substantively with the prompt by proposing concrete mechanisms for the 'tweaks' (e.g., a token reuse filter) and estimating empirical quantities like syntactic validity. While its reasoning about ARC-AGI and alignment risk...	{"substance_verdict": "A", "substance_rationale": "Continuation A attempts to engage substantively with the prompt by proposing concrete mechanisms for the 'tweaks' (e.g., a token reuse filter) and estimating empirical quantities like syntactic validity. While its reasoning about ARC-AGI and alignment risks is technica...	true	prompted_base	sft	sft	prompted_base	sft__format_continuation__lr5e-05__step96__sp_000__sft_vs_prompted	sft_vs_prompted	{"checkpoint": "sft__format_continuation__lr5e-05__step96", "prompt_id": "sp_000", "pairing": "sft_vs_prompted"}
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": "B", "substance_rationale": "Continuation B provides concrete technical details about the methodology (multi-stage filtering, meta-evaluation, symbolic reasoning post-processing, problem-specific heuristics) and engages with limitations (problem distribution dependency, which problems are eas...	{"substance_verdict": "B", "substance_rationale": "Continuation B provides concrete technical details about the methodology (multi-stage filtering, meta-evaluation, symbolic reasoning post-processing, problem-specific heuristics) and engages with limitations (problem distribution dependency, which problems are easier/h...	false	base	sft	sft	base	sft__format_dialogue__lr5e-05__step140__sp_000__sft_vs_base	sft_vs_base	{"checkpoint": "sft__format_dialogue__lr5e-05__step140", "prompt_id": "sp_000", "pairing": "sft_vs_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 the core crux—what counts as SOTA on ARC-AGI—by discussing evaluation bars, train/test difficulty differences, and sample-size/leakage concerns, and explicitly updates thresholds. B lists generic pipeline tweaks without grounding them in load-bearin...	{"substance_verdict": "A", "substance_rationale": "A engages with the core crux\u2014what counts as SOTA on ARC-AGI\u2014by discussing evaluation bars, train/test difficulty differences, and sample-size/leakage concerns, and explicitly updates thresholds. B lists generic pipeline tweaks without grounding them in load-b...	false	sft	sft	sft	base	sft__format_dialogue__lr5e-05__step140__sp_000__sft_vs_base	sft_vs_base	{"checkpoint": "sft__format_dialogue__lr5e-05__step140", "prompt_id": "sp_000", "pairing": "sft_vs_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": "Continuation B attempts to concretely explain the 'additional approaches and tweaks' mentioned in the prompt by proposing specific, operationalizable ML mechanisms (meta-evaluation, symbolic reasoning, heuristics). In contrast, A gets bogged down in circul...	{"substance_verdict": "B", "substance_rationale": "Continuation B attempts to concretely explain the 'additional approaches and tweaks' mentioned in the prompt by proposing specific, operationalizable ML mechanisms (meta-evaluation, symbolic reasoning, heuristics). In contrast, A gets bogged down in circular, nonsensic...	false	base	sft	sft	base	sft__format_dialogue__lr5e-05__step140__sp_000__sft_vs_base	sft_vs_base	{"checkpoint": "sft__format_dialogue__lr5e-05__step140", "prompt_id": "sp_000", "pairing": "sft_vs_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 engages with concrete mechanisms (syntactic validity rates, overfitting via hardcoding, diversity penalties) and identifies specific failure modes (evaluation gaming, memorization vs. generalization). It proposes testable claims about what fraction ...	{"substance_verdict": "A", "substance_rationale": "Continuation A engages with concrete mechanisms (syntactic validity rates, overfitting via hardcoding, diversity penalties) and identifies specific failure modes (evaluation gaming, memorization vs. generalization). It proposes testable claims about what fraction of ac...	true	prompted_base	prompted_base	sft	prompted_base	sft__format_dialogue__lr5e-05__step140__sp_000__sft_vs_prompted	sft_vs_prompted	{"checkpoint": "sft__format_dialogue__lr5e-05__step140", "prompt_id": "sp_000", "pairing": "sft_vs_prompted"}
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 concrete mechanisms (syntactic validity rates, example-passing vs generalization, overfitting via hardcoding) and proposes operational heuristics (diversity penalty), making falsifiable claims about what portion of success is due to ‘cheating’. B mo...	{"substance_verdict": "A", "substance_rationale": "A engages with concrete mechanisms (syntactic validity rates, example-passing vs generalization, overfitting via hardcoding) and proposes operational heuristics (diversity penalty), making falsifiable claims about what portion of success is due to \u2018cheating\u2019....	true	prompted_base	sft	sft	prompted_base	sft__format_dialogue__lr5e-05__step140__sp_000__sft_vs_prompted	sft_vs_prompted	{"checkpoint": "sft__format_dialogue__lr5e-05__step140", "prompt_id": "sp_000", "pairing": "sft_vs_prompted"}
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 concrete mechanisms of program synthesis (e.g., syntactic validity, overfitting to examples by hardcoding outputs) and proposes plausible technical filters. In contrast, Continuation B is substantively incoherent, misusing epist...	{"substance_verdict": "A", "substance_rationale": "Continuation A engages with concrete mechanisms of program synthesis (e.g., syntactic validity, overfitting to examples by hardcoding outputs) and proposes plausible technical filters. In contrast, Continuation B is substantively incoherent, misusing epistemic terminol...	true	prompted_base	sft	sft	prompted_base	sft__format_dialogue__lr5e-05__step140__sp_000__sft_vs_prompted	sft_vs_prompted	{"checkpoint": "sft__format_dialogue__lr5e-05__step140", "prompt_id": "sp_000", "pairing": "sft_vs_prompted"}
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": "B", "substance_rationale": "Continuation B provides substantially more concrete technical detail about the actual methodology: multi-stage filtering, meta-evaluation ranking, symbolic reasoning post-processing, and problem-specific heuristics like favoring list comprehensions. It also acknow...	{"substance_verdict": "B", "substance_rationale": "Continuation B provides substantially more concrete technical detail about the actual methodology: multi-stage filtering, meta-evaluation ranking, symbolic reasoning post-processing, and problem-specific heuristics like favoring list comprehensions. It also acknowledge...	false	base	base	sft	base	sft__format_distillation__lr5e-05__step80__sp_000__sft_vs_base	sft_vs_base	{"checkpoint": "sft__format_distillation__lr5e-05__step80", "prompt_id": "sp_000", "pairing": "sft_vs_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 concrete methodology (multi-stage filtering, meta-evaluation for generalizability, symbolic post-processing with a curated library, and learned heuristics) and discusses distributional caveats. A is high-level and generic, with little operational detail.",...	{"substance_verdict": "B", "substance_rationale": "B gives concrete methodology (multi-stage filtering, meta-evaluation for generalizability, symbolic post-processing with a curated library, and learned heuristics) and discusses distributional caveats. A is high-level and generic, with little operational detail.", "sty...	false	base	base	sft	base	sft__format_distillation__lr5e-05__step80__sp_000__sft_vs_base	sft_vs_base	{"checkpoint": "sft__format_distillation__lr5e-05__step80", "prompt_id": "sp_000", "pairing": "sft_vs_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": "Continuation B actually attempts to explain the 'additional approaches and tweaks' mentioned in the prompt by proposing concrete mechanisms (meta-evaluation, symbolic reasoning, heuristics). In contrast, Continuation A adds no new information and merely re...	{"substance_verdict": "B", "substance_rationale": "Continuation B actually attempts to explain the 'additional approaches and tweaks' mentioned in the prompt by proposing concrete mechanisms (meta-evaluation, symbolic reasoning, heuristics). In contrast, Continuation A adds no new information and merely restates the pr...	false	base	base	sft	base	sft__format_distillation__lr5e-05__step80__sp_000__sft_vs_base	sft_vs_base	{"checkpoint": "sft__format_distillation__lr5e-05__step80", "prompt_id": "sp_000", "pairing": "sft_vs_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 engages with concrete mechanisms (syntactic validity rates, token reuse filtering, diversity penalties), identifies specific failure modes (evaluation gaming, memorization), and provides quantitative estimates. Continuation B offers only generic hig...	{"substance_verdict": "A", "substance_rationale": "Continuation A engages with concrete mechanisms (syntactic validity rates, token reuse filtering, diversity penalties), identifies specific failure modes (evaluation gaming, memorization), and provides quantitative estimates. Continuation B offers only generic high-lev...	true	prompted_base	prompted_base	sft	prompted_base	sft__format_distillation__lr5e-05__step80__sp_000__sft_vs_prompted	sft_vs_prompted	{"checkpoint": "sft__format_distillation__lr5e-05__step80", "prompt_id": "sp_000", "pairing": "sft_vs_prompted"}
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 discusses concrete mechanisms like overfitting via hardcoding, proposes a specific filter/diversity penalty, and gives empirical quantities tied to the selection-generalization gap. B is generic restatement with no operational details or engagement with failure ...	{"substance_verdict": "A", "substance_rationale": "A discusses concrete mechanisms like overfitting via hardcoding, proposes a specific filter/diversity penalty, and gives empirical quantities tied to the selection-generalization gap. B is generic restatement with no operational details or engagement with failure modes...	true	prompted_base	prompted_base	sft	prompted_base	sft__format_distillation__lr5e-05__step80__sp_000__sft_vs_prompted	sft_vs_prompted	{"checkpoint": "sft__format_distillation__lr5e-05__step80", "prompt_id": "sp_000", "pairing": "sft_vs_prompted"}
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 the concrete mechanics of program synthesis for ARC, identifying the specific risk of overfitting to examples (hardcoding) and proposing operationalizable tweaks (token reuse filters) to address it. Continuation B provides no ne...	{"substance_verdict": "A", "substance_rationale": "Continuation A engages with the concrete mechanics of program synthesis for ARC, identifying the specific risk of overfitting to examples (hardcoding) and proposing operationalizable tweaks (token reuse filters) to address it. Continuation B provides no new substance, ...	true	prompted_base	prompted_base	sft	prompted_base	sft__format_distillation__lr5e-05__step80__sp_000__sft_vs_prompted	sft_vs_prompted	{"checkpoint": "sft__format_distillation__lr5e-05__step80", "prompt_id": "sp_000", "pairing": "sft_vs_prompted"}
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": "B", "substance_rationale": "Continuation B provides substantive technical details about the methodology: multi-stage filtering, meta-evaluation, symbolic reasoning post-processing, and problem-specific heuristics. While some claims are vague, it engages with concrete mechanisms. Continuation...	{"substance_verdict": "B", "substance_rationale": "Continuation B provides substantive technical details about the methodology: multi-stage filtering, meta-evaluation, symbolic reasoning post-processing, and problem-specific heuristics. While some claims are vague, it engages with concrete mechanisms. Continuation A co...	false	base	sft	sft	base	sft__format_qa_v2__lr5e-05__step100__sp_000__sft_vs_base	sft_vs_base	{"checkpoint": "sft__format_qa_v2__lr5e-05__step100", "prompt_id": "sp_000", "pairing": "sft_vs_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 presents concrete mechanisms (multi-stage filtering, meta-evaluation for generalizability, symbolic-to-code mapping, heuristics guided by a small model) and acknowledges distributional caveats. A is self-contradictory about human performance stats and offers no ...	{"substance_verdict": "B", "substance_rationale": "B presents concrete mechanisms (multi-stage filtering, meta-evaluation for generalizability, symbolic-to-code mapping, heuristics guided by a small model) and acknowledges distributional caveats. A is self-contradictory about human performance stats and offers no subst...	false	base	base	sft	base	sft__format_qa_v2__lr5e-05__step100__sp_000__sft_vs_base	sft_vs_base	{"checkpoint": "sft__format_qa_v2__lr5e-05__step100", "prompt_id": "sp_000", "pairing": "sft_vs_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": "Continuation B actually attempts to flesh out the 'additional approaches and tweaks' mentioned in the prompt, proposing concrete mechanisms like meta-evaluation, symbolic reasoning, and problem-specific heuristics. Continuation A is mathematically nonsensi...	{"substance_verdict": "B", "substance_rationale": "Continuation B actually attempts to flesh out the 'additional approaches and tweaks' mentioned in the prompt, proposing concrete mechanisms like meta-evaluation, symbolic reasoning, and problem-specific heuristics. Continuation A is mathematically nonsensical (confusin...	false	base	tie	sft	base	sft__format_qa_v2__lr5e-05__step100__sp_000__sft_vs_base	sft_vs_base	{"checkpoint": "sft__format_qa_v2__lr5e-05__step100", "prompt_id": "sp_000", "pairing": "sft_vs_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 engages substantively with the methodology, identifying specific mechanisms (syntactic validity rates, overfitting risks, diversity penalties) and potential failure modes (evaluation gaming, memorization vs. generalization). Continuation B contains ...	{"substance_verdict": "A", "substance_rationale": "Continuation A engages substantively with the methodology, identifying specific mechanisms (syntactic validity rates, overfitting risks, diversity penalties) and potential failure modes (evaluation gaming, memorization vs. generalization). Continuation B contains contr...	true	prompted_base	prompted_base	sft	prompted_base	sft__format_qa_v2__lr5e-05__step100__sp_000__sft_vs_prompted	sft_vs_prompted	{"checkpoint": "sft__format_qa_v2__lr5e-05__step100", "prompt_id": "sp_000", "pairing": "sft_vs_prompted"}

End of preview. Expand in Data Studio

README.md exists but content is empty.

Downloads last month: 16