muradil211/AetherSearch_DPO

🔭 AetherSearch DPO

Preference pairs for reasoning, retrieval, and evidence-grounded answers

🏠 Project · 🧠 DPO Model · 🧪 Training Code · 🎓 SFT Data · 🤖 SFT Model

Dataset overview

AetherSearch DPO contains 2,126 preference pairs for training an agentic-search policy after supervised fine-tuning. Every row provides one shared prompt, a preferred assistant continuation, and a non-preferred continuation.

This repository is deliberately train-only. It contains no benchmark results, run logs, model checkpoints, or held-out result files.

The released muradil211/AetherSearch_DPO checkpoint was trained in one DPO stage over all 2,126 pairs in this exact release, using the public AetherSearch/dpo training code on a separate server.

Release at a glance

Item	Value
Preference pairs	2,126
Unique normalized questions	2,126
Identical chosen/rejected pairs	0
Split	train only
Ordering	Deterministic seeded hash shuffle
Shuffle seed	42
Training file	train.jsonl

Source composition

Source dataset	Pairs	Share
TriviaQA	1,445	67.97%
MuSiQue	410	19.28%
Natural Questions	130	6.11%
WebQuestions	87	4.09%
2WikiMultiHopQA	54	2.54%
Total	2,126	100.00%

Preference composition

Pair type	Pairs
answer_hard_negative	1,166
query_hard_negative	289
true_full_trajectory_preference	235
insufficient_information_continue_search	173
evidence_misread_negative	94
premature_answer_negative	93
multi_hop_decomposition_negative	30
regression_protection_pair	28
query_refinement_negative	18
Total	2,126

Public schema

Every JSONL row contains exactly these eight fields, in this order:

Field	Type	Meaning
id	string	Canonical six-digit public identifier
question	string	Original question
source_dataset	string	Question-source dataset
answers	list[string]	Accepted answer aliases
pair_type	string	Preference/error category
prompt_text	string	Shared system, user, and trajectory context
chosen	string	Preferred assistant continuation
rejected	string	Non-preferred assistant continuation

The DPO training unit is:

(prompt_text, chosen, rejected)

chosen and rejected are continuations and do not duplicate prompt_text. Some full-trajectory continuations contain environment-provided <information>...</information> spans. A trainer reproducing the AetherSearch loss contract should exclude those environment tokens from preference loss.

Curation and integrity

Before publication:

1. Questions overlapping the referenced Search-R1 train/test sets or the AetherSearch SFT release were removed using normalized-question matching.
2. Only one preference pair was retained per normalized question.
3. Internal paths, stage labels, rollout filenames, and private source IDs were removed from the public schema.
4. Rows were globally reordered with a deterministic seeded hash shuffle, then assigned IDs 000001 through 002126.
5. Three spacing/punctuation-only normalizations standardized factual entity designations; preference meaning and labels were unchanged.

The release audit verifies 2,126 unique IDs, 2,126 unique normalized questions, 2,126 non-empty chosen continuations, 2,126 non-empty rejected continuations, and zero identical pairs. Full distributions and character-length summaries are recorded in dataset_manifest.json.

Load with Datasets

from datasets import load_dataset

dataset = load_dataset(
    "muradil211/AetherSearch_DPO",
    split="train",
)

print(dataset.column_names)
print(len(dataset))  # 2126

Files

File	Purpose
train.jsonl	Canonical preference-training split
dataset_manifest.json	Schema, distributions, curation, and integrity audit
checksums.sha256	Release-file integrity checksums
ATTRIBUTION.md	Source attribution and rights-status record
assets/aethersearch-mark.svg	AetherSearch release mark

Limitations

• Preference labels reflect curated negative construction and trajectory selection; they are not human preference votes for every row.
• Retrieved passages can contain incomplete, outdated, or incorrect evidence.
• The dataset is intended for preference training and is not a benchmark.
• No downstream quality or safety result is claimed by this data release.

Rights and attribution

No blanket license is asserted for the combined release. Review ATTRIBUTION.md and the terms of every upstream source before redistribution or downstream publication.

Checksums

After downloading the repository, verify the release with:

sha256sum -c checksums.sha256