datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
multireward-grpo-gsm8k-rewards-qwen2.5-7b
Multi-Reward GRPO — GSM8K Rewards (Qwen2.5-7B-Instruct)
Raw rollout-level reward observations from the empirical Section of
"Conditioned Multi-Reward Advantage Estimation: A Finite-Sample Analysis".
This is the data that produced the headline Theorem 3 (correlation-dependent
MSE floor) and Proposition 4 (sign-changing conditioning bias) figures on
real LLM rollouts. Each rollout was sampled from Qwen/Qwen2.5-7B-Instruct on
GSM8K test prompts at temperature 0.7.
What's in… See the full description on the dataset page: https://huggingface.co/datasets/eagle0504/multireward-grpo-gsm8k-rewards-qwen2.5-7b.reddit-posts-summarization-grpo
GRPO Summarization Eval Rollouts
Evaluation artifacts for all GRPO summarization checkpoints from smolcluster — a distributed GRPO training framework for Apple Silicon Mac clusters.
Two base models were fine-tuned across two training strategies and six reward configurations each, then evaluated on 200 examples from the mlabonne/smoltldr test split.
Judge: gpt-5-mini-2025-08-07 · Framework: DeepEval G-Eval · Rounds: 5 averaged · Metrics (each 0–1): Faithfulness · Coverage ·… See the full description on the dataset page: https://huggingface.co/datasets/YuvrajSingh9886/reddit-posts-summarization-grpo.danish-json-grpo-v1
danish-json-grpo-v1
10,015 Danish prompts for schema-directed JSON generation, built for GRPO training with a deterministic verifier (parse + key-set match + optional grounding penalty).
Task types
task_type
share
shape
extract
42%
Danish passage + schema → JSON grounded in passage
generate
26%
"Give me JSON for X with fields Y" (values open-ended)
rewrite
22%
Bullet list / semicolon-separated data → JSON with same info
fill_template
10%
JSON… See the full description on the dataset page: https://huggingface.co/datasets/jensjepsen/danish-json-grpo-v1.caliber-extension-gemma4-e2b-grpo-rollouts
CALIBER Extension — Gemma4-E2B GRPO Rollouts
Training rollouts from matched GRPO arms on google/gemma-4-E2B-it
(new-prompt template, non-thinking, full bf16, max completion 1500, 150 steps).
Subsets
subset
arm
τ
prior
rows
mean reward_total
accuracy
full schema
caliber
vanilla CALIBER
0.0
—
1600
2.298
0.514
0.664
mink
Min-K% prior
1.0
mink_0.2
4800
2.506
0.520
0.680
minkpp
Min-K++% prior
1.0
minkpp_0.2
4800
2.637
0.541
0.726
Load:
from datasets… See the full description on the dataset page: https://huggingface.co/datasets/dmnsh/caliber-extension-gemma4-e2b-grpo-rollouts.arabic-rag-chat-grpo-5K
Arabic multi-turn RAG conversations — GRPO pool (5,259 conversations)
The reinforcement-learning half of
oddadmix/arabic-rag-chat-30K:
same generator, same validator, same schema, disjoint companies. It exists
so GRPO explores fresh knowledge bases instead of taking a second pass over
material the SFT already memorised.
conversations
turns
companies
this pool
5,259
14,018
309
Company-disjointness is exact and verified: this pool shares zero
company_id values… See the full description on the dataset page: https://huggingface.co/datasets/oddadmix/arabic-rag-chat-grpo-5K.debate-grpo-group-a
Debate GRPO Group A - TACTIC_SELECT
Training data for debate model GRPO fine-tuning (Group A: TACTIC_SELECT calls).
Files
File
Description
Rows
group_a_rescored_v2_with_logps.parquet
Training format (one row per response) with precomputed logprobs
1,993
group_a_flat_rescored_v2.parquet
Flat format with RESPONSE_1-6 columns per call
520
Training Format Columns
Column
Description
debate_id
Unique debate identifier
call_id… See the full description on the dataset page: https://huggingface.co/datasets/debaterhub/debate-grpo-group-a.ipda-grpo-training-data
IPDA GRPO Training Data
Training data for GRPO (Group Relative Policy Optimization) on IPDA debate tasks.
Dataset Description
Contains scored debate speech samples used for GRPO training iterations. Each sample includes:
Input prompt (debate context)
Generated response (speech)
Rubric scores from debate judge
Log probabilities for policy optimization
Files
File
Description
Samples
group_c_grpo.parquet
Group C (warrant/clash) training data
~3K… See the full description on the dataset page: https://huggingface.co/datasets/dgonier/ipda-grpo-training-data.multireward-grpo-fintech-customer-comms
Multi-Reward GRPO — Synthetic Fintech Customer Communications
Synthetic multi-turn customer-service conversations for a fictional bank
("Bank of XYZ"), generated for the empirical Section of "Conditioned
Multi-Reward Advantage Estimation: A Finite-Sample Analysis".
Each conversation ends with m parallel sampled bot replies, each scored
on three verifiable reward channels designed for fintech customer service.
This is the multi-reward GRPO group structure on a real generation… See the full description on the dataset page: https://huggingface.co/datasets/eagle0504/multireward-grpo-fintech-customer-comms.danish-if-grpo-combined-v3
danish-if-grpo-combined-v3
10,000 Danish instruction-following prompts for GRPO reward training,
built by rewriting danish-instruction-following-v4 prompts with a
sampled mix of our 46 Danish constraints + ~24 Google IFEval-schema
constraints.
What's new vs v2
Structural fixes (see scripts/constraint_compat.py on
github):
0% impossible constraint combos (v2 had 15.2%; e.g. constrained_response
paired with 20+ content-shape rules that can't co-satisfy). Enforced… See the full description on the dataset page: https://huggingface.co/datasets/jensjepsen/danish-if-grpo-combined-v3.nemo-grpo-weak3-from084-prompts
Nemo Weak-3 GRPO Prompt Dataset
This dataset is a clean GRPO/RLVR prompt set for the three weak Nemotron
challenge types identified after the 0.84 SDPO adapter diagnostics:
bit_manipulation, unit_conversion, and gravity.
The training rows are intentionally modeled as:
prompt x + gold answer r + verifier/reward spec
There are no source CoT traces, teacher completions, SDPO samples, RLSD
privileged traces, or eval predictions in the training split. GRPO should
sample completions… See the full description on the dataset page: https://huggingface.co/datasets/dvyomkesh/nemo-grpo-weak3-from084-prompts.azure-advisor-grpo-benchmark
Azure Advisor GRPO Benchmark Dataset
Evaluation benchmark for measuring the quality of Azure Advisor recommendation generation, used for GRPO (Group Relative Policy Optimization) training and model evaluation.
Dataset Description
This dataset contains 106 evaluation examples with ground truth labels, designed to score model outputs across 5 reward dimensions.
Purpose
During GRPO training: Score generated recommendations to select high-reward samples
Model… See the full description on the dataset page: https://huggingface.co/datasets/thegovind/azure-advisor-grpo-benchmark.deepseek_grpo_correct_6144
DeepSeek GRPO Correct 6144
Filtered GRPO training subset generated from deepseek-reasoner math generations.
Files
train.jsonl: filtered training examples with prompt, solution, dataset_index, and DeepSeek metadata.
metadata.json: filtering metadata.
Filter
Rows are kept when the raw generation is successful, stopped, correct, deduplicated by dataset_index, and has usage_total_tokens <= 6144.
Summary
Rows: 7576
Max total tokens: 6144
Source raw… See the full description on the dataset page: https://huggingface.co/datasets/igreck/deepseek_grpo_correct_6144.deepseek_grpo_correct_8192
DeepSeek GRPO Correct 8192
Filtered GRPO training subset generated from deepseek-reasoner math generations.
Files
train.jsonl: filtered training examples with prompt, solution, dataset_index, and DeepSeek metadata.
metadata.json: filtering metadata.
Filter
Rows are kept when the raw generation is successful, stopped, correct, deduplicated by dataset_index, and has usage_total_tokens <= 8192.
Summary
Rows: 9002
Max total tokens: 8192
Source raw… See the full description on the dataset page: https://huggingface.co/datasets/igreck/deepseek_grpo_correct_8192.multireward-grpo-gsm8k-rewards
Multi-Reward GRPO — GSM8K Rewards (Qwen2.5-1.5B-Instruct)
Raw rollout-level reward observations from the empirical Section of
"Conditioned Multi-Reward Advantage Estimation: A Finite-Sample Analysis".
This is the data that produced the headline Theorem 3 (correlation-dependent
MSE floor) and Proposition 4 (sign-changing conditioning bias) figures on
real LLM rollouts.
What's in here
For each of 150 GSM8K test prompts, we sampled 16 independent seeds × 32
rollouts… See the full description on the dataset page: https://huggingface.co/datasets/eagle0504/multireward-grpo-gsm8k-rewards.deepseek_grpo_correct_4096
DeepSeek GRPO Correct 4096
Filtered GRPO training subset generated from deepseek-reasoner math generations.
Files
train.jsonl: filtered training examples with prompt, solution, dataset_index, and DeepSeek metadata.
metadata.json: filtering metadata.
Filter
Rows are kept when the raw generation is successful, stopped, correct, deduplicated by dataset_index, and has usage_total_tokens <= 4096.
Summary
Rows: 5502
Max total tokens: 4096
Source raw… See the full description on the dataset page: https://huggingface.co/datasets/igreck/deepseek_grpo_correct_4096.RLVE-Test-Qwen3-1.7B-GRPO-step70-Pass8
RLVE test eval — GRPO step70 (pass@8)
Evaluation rollouts on the RLVE test split.
Model: grpo_train_Qwen3-1.7B-SFT-rlve-20K-1epoch (GRPO, step 70)
Source prompts: RLVE test split — 180 questions (RLVE-Eval Gym environments)
Sampling: 8 samples/question (pass@8) = 1440 records, temperature 0.7,
max 16384 new tokens
Rewards: inline RLVE-Eval Gym verifier score (continuous, in [-1, 1]).
Record-level accuracy (reward>0): 216 / 1440 = 15.0%, mean reward -0.677
pass@8 (>=1 of 8… See the full description on the dataset page: https://huggingface.co/datasets/CL-From-Nothing/RLVE-Test-Qwen3-1.7B-GRPO-step70-Pass8.deepseek_grpo_correct_2048
DeepSeek GRPO Correct 2048
Filtered GRPO training subset generated from deepseek-reasoner math generations.
Files
train.jsonl: filtered training examples with prompt, solution, dataset_index, and DeepSeek metadata.
metadata.json: filtering metadata.
Filter
Rows are kept when the raw generation is successful, stopped, correct, deduplicated by dataset_index, and has usage_total_tokens <= 2048.
Summary
Rows: 2153
Max total tokens: 2048
Source raw… See the full description on the dataset page: https://huggingface.co/datasets/igreck/deepseek_grpo_correct_2048.GRPO-Fine-Tuned
GSM8K GRPO Dataset for Qwen3-4B Post-Training
A clean, GRPO-ready dataset derived from GSM8K for post-training the Qwen3-4B base model using Group Relative Policy Optimization (GRPO).
Dataset Purpose
This dataset is designed for the GRPO stage of post-training, where the model learns to produce correct mathematical reasoning through reward-based optimization. The key design principles are:
Verifiable answers: Every example has a single, unambiguous numeric answer
Clean… See the full description on the dataset page: https://huggingface.co/datasets/TeamClaude/GRPO-Fine-Tuned.debate-multi-trial-grpo
Debate Multi-Trial GRPO Training Data
Training data for offline GRPO (Group Relative Policy Optimization) on IPDA debate generation.
Dataset Structure
Each row represents one pipeline call with 4 response variants:
RESPONSE_1_* through RESPONSE_4_*: Different generations at varying temperatures
*_SCORE: Quality score (0.0-1.0) from Haiku evaluator
chosen_index: Index of highest-scoring response
rejected_index: Index of lowest-scoring response
Statistics… See the full description on the dataset page: https://huggingface.co/datasets/debaterhub/debate-multi-trial-grpo.ipda_grpo_multi_trial_thinking_tactics
Debate Multi-Trial GRPO Test Data (with Thinking Frameworks)
TEST DATASET - Single debate for review before scaling.
Training data for offline GRPO (Group Relative Policy Optimization) on IPDA debate generation,
with integrated thinking framework injection.
What's New: Thinking Frameworks
Each prompt includes structured thinking instructions (mnemonics) that guide the model's reasoning:
Call Type
Mnemonic
Purpose
TACTIC_SELECT
JAM
Judge-Attack-Momentum Analysis… See the full description on the dataset page: https://huggingface.co/datasets/debaterhub/ipda_grpo_multi_trial_thinking_tactics.ipda-grpo-dataset-iter3-feb-12
IPDA GRPO Dataset — Iteration 3 (Feb 12, 2026)
GRPO (Group Relative Policy Optimization) training dataset for IPDA (International Public Debate Association) debate speech generation.
Dataset Structure
2,988 unique prompts | 11,425 scored trials | Score avg: 0.700 (0-1 scale)
Each row represents a unique debate pipeline prompt with up to 6 trial responses:
Column
Description
prompt_hash
SHA256[:16] of prompt text
prompt
Full pipeline prompt
speech_type
AC… See the full description on the dataset page: https://huggingface.co/datasets/dgonier/ipda-grpo-dataset-iter3-feb-12.Evaluation_GRPO