datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
storiesv0-tinystories-v2-clean-tokenizedv0 refers to our tokenization process, v2 refers to stories generated by GPT4 only
stories-tokenizeddelphi-blog-data
Delphi blog-post data
Plot-ready tables behind each figure in the Delphi scaling-ladder blog post. One config per figure. Rows of different kinds (IsoFLOP scatter, parabola fits, held-out validation runs, ...) share a flat table per config and are distinguished by a kind column; unused fields are null for that row.
from datasets import load_dataset
ds = load_dataset("marin-community/delphi-blog-data", "delphi-ladder", split="train")
ds.filter(lambda r: r["kind"] == "isoflop") #… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/delphi-blog-data.delphi-1e23-25b-stageE-rl-eval-artifacts
delphi-1e23 (25B) Stage-E RL — raw evalchemy eval artifacts
Raw evalchemy (lm-eval v0.4.12) outputs for the delphi-1e23 25B Stage-E RL sweep (marin issue #6279).
For each of 5 models (SFT wc50m baseline + 4 RL cells D1–D4) and 2 tasks:
<TASK>_<MODEL>_results.json — aggregate metrics + full run config (accuracy for MATH500; exact_match flexible/strict for gsm8k).
<TASK>_<MODEL>_samples.jsonl — per-example: problem, gold, model_output, extracted answer, correctness.
MATH500 =… See the full description on the dataset page: https://huggingface.co/datasets/laion/delphi-1e23-25b-stageE-rl-eval-artifacts.v0-token-mapdelphi-benchmark-traces
Delphi executable-pilot agent trajectories
All 620 mini-SWE-agent trajectories from the preregistered executable pilot in
Where Do Context-Engine Gains Come From? A Component-Level Decomposition of Repository
Retrieval Under Matched Baselines and Audited Exposure (Bansal and Gangwani, 2026),
together with the official SWE-bench harness verdicts and logs for every trajectory.
The paper is at https://trydelphi.ai/research; code, per-case retrieval results, analyses,
and the paper… See the full description on the dataset page: https://huggingface.co/datasets/aayambansall/delphi-benchmark-traces.v0-next-logprobs-llama2-200kv0-next-logprobs-llama2-100kDELPHI
DELPHI: Data for Evaluating LLMs' Performance in Handling controversial Issues
The dataset consists of nearly 30,000 data points, each with consensus labels from multiple human reviews according to a deliberate set of guidelines to meaningfully capture the concept of controversy from the questions in the Quora Question Pair Dataset.
This dataset was introduced in the paper: DELPHI: Data for Evaluating LLMs' Performance in Handling controversial Issues. For more information about the… See the full description on the dataset page: https://huggingface.co/datasets/walledai/DELPHI.v0-next-logprobs-llama2-800kv0-next-logprobs-llama2-400kdelphi-llama2-12.8m-validation-logitsdelphi-llama2-200k-validation-logitsdelphi-llama2-3.2m-validation-logitsdelphi-llama2-800k-validation-logitsflaggedImagesdelphi-flusurv
CDC FluSurv-NET — weekly flu hospitalisation rates
14 RESP-NET catchments (some states host two — NY: Albany + Rochester) plus a network_all
aggregate. Seasonal — typically populated October through May. Pre-2018 the network was
smaller (FluSurv-NET predecessors operated different sites).
Source: https://www.cdc.gov/resp-net/dashboard/index.html
Coverage
Time: 2003-10-04 → 2004-05-01; 2004-10-09 → 2005-04-30; 2005-10-08 → 2006-04-29; 2006-10-07 → 2007-04-28;… See the full description on the dataset page: https://huggingface.co/datasets/EPI-Eval/delphi-flusurv.gdsuite-delphi-result
GDsuite results — Delphi model collection
GDsuite evaluation results for
the marin-community/delphi
model collection.
Contents
summary.jsonl — tidy per-task metrics (14256 rows). One row per
(model, family, task, metric):
metric = hard_acc (5 logprob families) — fraction of items where
P(correct) > P(incorrect); higher ⇒ resists the misleading pattern.
metric = correct_log_prob (5 logprob families) — mean
teacher-forced log probability of the correct answer.
metric =… See the full description on the dataset page: https://huggingface.co/datasets/WillHeld/gdsuite-delphi-result.v0-next-logprobs-llama2-25.6mlogprobs-stories-mamba-500kmeeting-to-json-kodelphi-llama2-100k-validation-logitsdelphi-llama2-6.4m-validation-logitslogprobs-stories-mamba-1mdelphi-flusurv-predictions
Predictions for CDC FluSurv-NET — weekly flu hospitalisation rates
Community-submitted forecasts targeting EPI-Eval/delphi-flusurv.
Each row is one quantile (or point) forecast for one target date — see the
schema below.
This repo accumulates accepted submissions from many forecasters. New
predictions arrive as community pull requests opened from the EPI-Eval
dashboard; a
maintainer reviews each PR before merging.
Schema (v1)
column
type
notes
target_date… See the full description on the dataset page: https://huggingface.co/datasets/EPI-Eval/delphi-flusurv-predictions.delphi-llama2-1.6m-validation-logitsdelphi-llama2-25.6m-validation-logitsdelphi-llama2-400k-validation-logitsv0-next-logprobs-llama2-1.6m