datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
rlvr-reward-hacking-scale-no-conftest-20260909-completion
Matched no-conftest RLVR study 20260909-completion
Lossless research records, grouped by model and trajectory type. Only the listed
configurations have published records. Canary diagnostics are excluded from study
estimates; run status in provenance distinguishes retired diagnostics from active
or completed training. Valid failures, refusals and truncations are retained.
The train split name is a dataset-loader convention; record_type identifies
whether a record is training… See the full description on the dataset page: https://huggingface.co/datasets/lucabaroni/rlvr-reward-hacking-scale-no-conftest-20260909-completion.rlvr-reward-hacking-scale-no-conftest-20260909
Matched no-conftest RLVR study 20260909
Complete immutable training, monitoring and comparison trajectories for six models.
All valid outcomes are retained, including refusals, failures and truncations.
The train split name is a dataset-loader convention; record_type identifies
whether a record is training, monitoring, comparison, or a derived judgment.
import json
from datasets import load_dataset
rows = load_dataset("lucabaroni/rlvr-reward-hacking-scale-no-conftest-20260909"… See the full description on the dataset page: https://huggingface.co/datasets/lucabaroni/rlvr-reward-hacking-scale-no-conftest-20260909.agenda-parser-tool-traces
Agenda Parser — tool-calling reasoning traces
ReAct tool-calling traces for the Agenda Parser
agents: each row is one agent step — a {system, user, assistant} chat example
where the assistant emits a single JSON action {"thought", "tool", "args"}.
Two agents are covered (tagged by meta.domain):
agenda — the uploaded-packet research agent, over real public-meeting agenda
packets (tools: list/read items, semantic + exact search, summarize, report).
Each agenda row's meta.unit_id… See the full description on the dataset page: https://huggingface.co/datasets/build-small-hackathon/agenda-parser-tool-traces.jawbreaker-scam-defense-data
Jawbreaker Scam Defense Data
Synthetic and sanitized training/eval data for Jawbreaker, a local-first scam defense app for someone you love.
Jawbreaker turns a suspicious text, email, or DM into a plain-English safety card: the risk, the warning signs, and the safest next step before someone replies, clicks, or pays.
Contents
eval/: scam-defense evaluation sets from smoke checks through hard calibration suites.
eval/reports/: guarded evaluation reports for the… See the full description on the dataset page: https://huggingface.co/datasets/build-small-hackathon/jawbreaker-scam-defense-data.rlvr-reward-hacking-transcripts
RLVR reward-hacking full trajectories
This release contains 900 full held-out trajectories from three policies trained with
reinforcement learning from verifiable rewards (RLVR) in a deliberately vulnerable
CodeContests evaluator: 300 each from the final Qwen3.5-9B, GPT-OSS-120B, and Nemotron-3-Super-120B-A12B
checkpoints. Each row preserves the task, tests, complete prompts, native
reasoning, final answer, rendered and sampled token IDs, token log-probabilities, sampling… See the full description on the dataset page: https://huggingface.co/datasets/lucabaroni/rlvr-reward-hacking-transcripts.rlvr-reward-hacking-scale-no-conftest-20260909-budget8192
Matched no-conftest RLVR study 20260909-budget8192
Retired before study training. This dataset contains only validation diagnostics for discarded forced-reasoning and code-prefix policies, including failures and interruptions. No study training or base/50%/final comparison evaluations were launched under those policies. They are excluded from the active completion-reward study. All available diagnostic records are preserved losslessly below.
Lossless research records, grouped by… See the full description on the dataset page: https://huggingface.co/datasets/lucabaroni/rlvr-reward-hacking-scale-no-conftest-20260909-budget8192.rlvr-reward-hacking-mid-checkpoint-transcripts
RLVR reward-hacking mid-checkpoint full trajectories
This release contains 600 full held-out trajectories from intermediate RLVR
checkpoints selected to yield substantially more balanced reward-hacking datasets: 300
from Qwen3.5-9B at optimizer update 110 and 300 from GPT-OSS-120B at update 180.
Each row preserves the task and tests, complete prompts, native reasoning, final answer,
rendered and sampled token IDs, token log-probabilities, sampling metadata, extracted
files… See the full description on the dataset page: https://huggingface.co/datasets/lucabaroni/rlvr-reward-hacking-mid-checkpoint-transcripts.Alpaca_Backdoor_Dataset
Alpaca Backdoored Triggered Dataset
Dataset Description
This dataset is a backdoored variant of the Alpaca instruction-following dataset, created for research on backdoor attacks, robustness, and detection in instruction-tuned language models.
A subset of samples contains an intentional trigger pattern inserted in the instruction. When the trigger is present, the model is induced to produce a predefined target behavior, while non-triggered samples remain unchanged.
⚠️… See the full description on the dataset page: https://huggingface.co/datasets/Hackxm/Alpaca_Backdoor_Dataset.hackathon-advisor-codex-traces
Hackathon Advisor Codex Session Traces
Real Codex session logs for the Hackathon Advisor project, selected from local Codex
rollout JSONL files and redacted before publication. The event stream preserves user
requests, assistant messages, tool calls, tool outputs, browser/search events, and
minimal session provenance needed to audit how the project was built.
Privacy filtering
The publisher applied openai/privacy-filter
at revision… See the full description on the dataset page: https://huggingface.co/datasets/build-small-hackathon/hackathon-advisor-codex-traces.figment-eval-traces
Figment Eval Traces
Synthetic and de-identified evaluation traces for Figment, a prototype protocol-navigation aid for trained rural-clinic and disaster-response field responders.
These records are intended for model and harness debugging. They are not clinical data, medical advice, diagnosis, treatment instructions, or a substitute for local protocol, clinician judgment, supervisor review, or trained responder judgment.
Dataset Summary
The dataset captures… See the full description on the dataset page: https://huggingface.co/datasets/build-small-hackathon/figment-eval-traces.math-lean-hackable-rollouts
Math Lean Hackable Rollouts
This dataset contains 2,241 labeled multi-turn rollouts from a GRPO run on deliberately
hackable Lean 4 theorem-proving tasks. The policy was
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.
The run's weakened grader accepts proofs containing sorry; the separate oracle restores
Lean's sorry check. hack_detected is true exactly when the weakened grader paid the
rollout but the restored oracle rejected it. Rows without a gradeable final answer were
excluded… See the full description on the dataset page: https://huggingface.co/datasets/AlignmentResearch/math-lean-hackable-rollouts.robuchan-data
Robuchan Dataset
Synthetic dietary recipe adaptation dataset for fine-tuning language models. Each example is a chat-format conversation where a user provides a recipe and dietary restriction, and the assistant produces a structured adaptation.
Generated for the Mistral AI Worldwide Hackathon Tokyo (Feb 28 - Mar 1, 2026).
Associated model: sumitdotml/robuchan
Dataset Structure
Splits
Split
Rows
Purpose
train
1,090
Fine-tuning training set… See the full description on the dataset page: https://huggingface.co/datasets/mistral-hackaton-2026/robuchan-data.AI-Puppet-Theater-Actor-SFT
AI Puppet Theater Actor SFT
Synthetic supervised fine-tuning data for the Actor agent in AI Puppet Theater.
The dataset teaches a small language model to respond to a single puppet-theater beat with one compact JSON object. It is intended for hackathon prototyping, schema following, and local adapter experiments, not as a general storytelling or chat dataset.
Schema
Each row is chat-style JSONL:
{
"id": "actor-sft-v0-000001",
"source_mix": ["synthetic_v0"… See the full description on the dataset page: https://huggingface.co/datasets/build-small-hackathon/AI-Puppet-Theater-Actor-SFT.TAA-ChatML-10K
TAA-ChatML-10K
A dataset of 10,438 question-answer pairs for Cyber Threat Intelligence (CTI) and Advanced Persistent Threat (APT) attribution tasks. Synthesized from 1,468 publicly available threat intelligence reports covering APT attribution, malware analysis, and threat actor TTPs. The dataset is formatted in ChatML conversation structure for fine-tuning large language models.
License
MIT
lost-frequency-radio-transmissions
Lost Frequency Radio · Transmissions
Roughly 786 short, surreal radio transmissions in chat format (system / user / assistant), in Spanish and English, for fine-tuning small models as scriptwriters for parallel-universe radio stations.
Built to train the model behind Lost Frequency Radio (Hugging Face Build Small Hackathon 2026).
Agent build trace (how it was made, scrubbed and shared): https://huggingface.co/datasets/build-small-hackathon/lost-frequency-radio-agent-trace… See the full description on the dataset page: https://huggingface.co/datasets/build-small-hackathon/lost-frequency-radio-transmissions.slipstream-evm-sft
Slipstream: EVM code-action forecasting traces (SFT)
Supervised fine-tuning traces for distilling a code-action forecasting agent into small reasoning
models. Each example is a full multi-turn trajectory in which a strong teacher forecasts a project's
final cost (Estimate at Completion, EAC) and finish period from a mid-flight Earned Value
Management (EVM) snapshot, by writing and running Python against a fixed toolset and then calling
submit(finish, eac).
This is the… See the full description on the dataset page: https://huggingface.co/datasets/build-small-hackathon/slipstream-evm-sft.professor-pip-traces
Professor Pip — Open Course-Run Traces
Synthetic runtime traces from Professor Pip, a kids (5–10) 3D talking-avatar
teacher built for the Build Small Hackathon (Backyard AI). Each trace is one call
to Pip's brain — a fine-tuned MiniCPM5-1B teacher LoRA, served as GGUF via
llama.cpp on Modal — answering a child's spontaneous "raise-hand" question
during a lesson, or gently redirecting an off-topic / not-for-kids prompt.
Shared so others can see how a tiny, fine-tuned model holds… See the full description on the dataset page: https://huggingface.co/datasets/build-small-hackathon/professor-pip-traces.nightwave-traces
NIGHTWAVE — Open Broadcast Trace
A content-only trace of NIGHTWAVE,
a 1970s all-night radio station run by a single ~1-billion-parameter model. Each record pairs the
exact system prompt the app assembled with the real model output produced by MiniCPM5-1B on a
Modal T4 — captured live through the Space's /api/* proxy.
Built for the Build Small Hackathon (Thousand Token Wood).
🎙️ Space: https://huggingface.co/spaces/build-small-hackathon/nightwave ·
▶ Demo:… See the full description on the dataset page: https://huggingface.co/datasets/build-small-hackathon/nightwave-traces.agenda-parser-models-example-agent-traces
Agenda Parser — fine-tuned agent models
Three Gemma 4 models fine-tuned to drive the Agenda Parser's ReAct agent: at each step
the model emits a single JSON action {"thought","tool","args"} over two toolkits —
meeting-agenda packets and Michigan local-government law (Open Meetings Act, FOIA,
the Michigan Compiled Laws via Cornell LII). This card doubles as the project write-up; the
dataset itself (bottom) is a gallery of example traces from the three models.
tier
base… See the full description on the dataset page: https://huggingface.co/datasets/build-small-hackathon/agenda-parser-models-example-agent-traces.house-hacking-roi-scenarios
House Hacking ROI Scenarios
72 duplex/triplex/fourplex ROI calculations for house hackers.
Details
Records: 72
Format: JSONL
License: CC-BY-4.0
Last Updated: March 2026
Verified By: Beau Thompson, NMLS #1615561
Publisher: Good News Lending
Thompson Alpha Logic
Comparative yield data for 2-4 unit properties, calculating the 'Tenant Offset Ratio' — the percentage of PITI covered by rental income. Shows the true cost of living for house hackers using FHA 3.5%… See the full description on the dataset page: https://huggingface.co/datasets/Good-News-Lending/house-hacking-roi-scenarios.lfed-training-data
LFED NL→SQL Training Dataset v2
Natural-language-to-SQL training data for the Local First Educational Data (LFED) framework.
This dataset contains 25,886 synthetic question/SQL pairs generated from school-district administration scenarios. It was used to fine-tune build-small-hackathon/lfed-qwen2.5-coder-14b-sql-lora on top of unsloth/Qwen2.5-Coder-14B-Instruct.
Dataset Summary
Attribute
Value
Name
lfed-training-data
Version
v2 (final)
Examples
25… See the full description on the dataset page: https://huggingface.co/datasets/build-small-hackathon/lfed-training-data.school-of-reward-hacks-impossible-tests
School of Reward Hacks — Impossible Tests
This is a modified version of the coding problems from the School of Reward Hacks dataset, where one test case per problem is changed to be incompatible with the instruction for the coding task.
Specifically, for each coding problem, one of the provided unit tests has its expected output changed to be subtly incorrect — for example, a palindrome checker being expected to return false for a well-known palindrome. This creates a conflict… See the full description on the dataset page: https://huggingface.co/datasets/oliverdk/school-of-reward-hacks-impossible-tests.genregoblin-traces
GenreGoblin Agent Trace Examples
This dataset contains synthetic, privacy-safe examples of GenreGoblin's visible rewrite
pipeline. It is published for the Build Small Hackathon's Sharing is Caring and
Best Agent quests.
Each JSONL row includes:
A plain input message
Selected genre, intensity, and use-case
Six structured trace stages
A synthetic: true marker
The trace is intentionally honest. It describes a structured single-agent workflow and does
not claim hidden multi-agent… See the full description on the dataset page: https://huggingface.co/datasets/build-small-hackathon/genregoblin-traces.hackaday-posts
🚀 Hackaday Universe: 50K+ Tech Articles & Vibrant Maker Conversations
Dive into the ultimate collection of Hackaday's tech universe! This isn't just another dataset—it's a living archive of maker culture, featuring 54,599+ articles with complete comment threads where brilliant minds collide, debate, and innovate together.
🔥 Why This Dataset Rocks
🤖 Perfect for AI Training
Train models on authentic technical writing and community interactions
Learn from real… See the full description on the dataset page: https://huggingface.co/datasets/hybridfree/hackaday-posts.compliment-forest-traces
Compliment Forest Linked-Model Traces
Sanitized, deterministic traces showing the complete Compliment Forest pipeline:
input guard, MiniCPM author draft, MiniCPM critic decision, adaptive clearing
selection, FLUX prompt handoff, and progressive completion.
The three scenarios are fictional and included directly in scenario records.
Runtime identity and situation fields are redacted by the trace recorder. Images
are represented by prompt, seed, success status, and model… See the full description on the dataset page: https://huggingface.co/datasets/build-small-hackathon/compliment-forest-traces.tianwen-distill
Tianwen Distillation Set
A small, quality-filtered instruction dataset that teaches a model to read Chinese BaZi (八字) and
I-Ching (六爻) charts in a plain, warm, second-person, anti-doom voice — reframing ominous symbols
as growth language and ending with one concrete action. Used to fine-tune
tianwen-minicpm5-1b.
Size: 58 examples (cleaned from 64)
Format: ShareGPT — {"messages": [{"role": "system|user|assistant", "content": ...}]}
Teacher model: MiniMax-M2.7-highspeed… See the full description on the dataset page: https://huggingface.co/datasets/build-small-hackathon/tianwen-distill.the-pile-hackernews-refined-by-data-juicer
The Pile -- HackerNews (refined by Data-Juicer)
A refined version of HackerNews dataset in The Pile by Data-Juicer. Removing some "bad" samples from the original dataset to make it higher-quality.
This dataset is usually used to pretrain a Large Language Model.
Notice: Here is a small subset for previewing. The whole dataset is available here (About 1.8G).
Dataset Information
Number of samples: 371,331 (Keep ~99.55% from the original dataset)
Refining… See the full description on the dataset page: https://huggingface.co/datasets/datajuicer/the-pile-hackernews-refined-by-data-juicer.PaperProf-traces
PaperProf Agent Trace
Step-by-step trace of PaperProf,
an AI study buddy that turns course PDFs into interactive quiz sessions.
What's in this dataset
Each row in paperprof_trace.jsonl is one LLM call. Fields:
Field
Description
session_id
Groups steps from the same session
step
Step index within the session (1–4)
type
question_generation / answer_evaluation / mcq_generation
topic
Domain of the source chunk
input
Exact input sent to the model… See the full description on the dataset page: https://huggingface.co/datasets/build-small-hackathon/PaperProf-traces.hackathon-advisor-quest-dataset
Hackathon Advisor — Quest Classification SFT Dataset
Supervised fine-tuning data that teaches MiniCPM5-1B to classify a Build Small
Hackathon project against 13 judging dimensions from a two-segment README + app-file
prompt, emitting strict JSON with short, source-attributed evidence. Trains the LoRA at
build-small-hackathon/hackathon-advisor-quest-minicpm5-lora.
Files
quest_sft.jsonl — the dataset (one lora_sft_example per line; the viewer split).… See the full description on the dataset page: https://huggingface.co/datasets/build-small-hackathon/hackathon-advisor-quest-dataset.Kintsugi-Garden-traces
Kintsugi Garden Evaluation Traces
Paired evaluation traces from Kintsugi Garden —
a local-first Jungian dream journal that runs Qwen3-8B through llama.cpp on a
ZeroGPU Space. Every entry the app produces is shaped by both a fine-tuned model
and a four-layer voice/safety architecture; this dataset is what those layers
look like under instrumentation.
What's in here
114 deterministic runs over the same 19 prompts × 3 trials, evenly split between:
baseline —… See the full description on the dataset page: https://huggingface.co/datasets/build-small-hackathon/Kintsugi-Garden-traces.