geyuxu/york-milestone-project-self-traing-loop-data

York Milestone Project Self-Training Loop — Phase 1 Preferences

Status and access restrictions

This repository contains the frozen historical Phase 1 preference corpus used by the York milestone self-training-loop project. It is not a Phase 2 leakage-safe training release.

Keep the repository private unless and until the upstream dataset licences, document-level redistribution rights, privacy/PII review, and generated-content policy review have all been completed. The license: other metadata is intentional: this staging release does not claim that one licence covers all included source passages or generated fields.

Contents

The published data are a mechanically converted copy of the final Phase 1 deduplicated preference universe:

Item	Count
Preference pairs	111,096
retrieval_helps	85,687
retrieval_hurts	25,409
Historical train IDs	99,986
Historical validation IDs	11,110
Duplicate pair_id values	0
Rows with null retrieved_documents	1,847

The source contributions recorded by the Phase 1 build are:

data_source	Unique pairs
expanded_mixed_500k	85,780
synthetic_data_expanded_mixed_supplement_001	14,747
previous_valid_11523	10,569

The expanded evidence pool used TriviaQA rc evidence, HotpotQA fullwiki contexts, and SQuAD train contexts. Human questions, human answers, and supporting-fact labels from those datasets were not used as preference labels; the preference records were generated and scored automatically. This distinction does not remove the need to review the redistribution terms of the underlying passages.

Repository layout

.
├── .gitattributes
├── README.md
├── MANIFEST.sha256
└── phase1
    ├── data
    │   ├── phase1-all-00000-of-00006.parquet
    │   ├── phase1-all-00001-of-00006.parquet
    │   ├── phase1-all-00002-of-00006.parquet
    │   ├── phase1-all-00003-of-00006.parquet
    │   ├── phase1-all-00004-of-00006.parquet
    │   └── phase1-all-00005-of-00006.parquet
    ├── split.json
    └── stats.json

phase1/data/*.parquet contains all 111,096 records in stable source order. phase1/split.json stores the historical Phase 1 membership as pair_id lists; the Parquet files themselves are not physically partitioned by the old train/validation assignment.

Example loading:

from datasets import load_dataset

all_pairs = load_dataset(
    "geyuxu/york-milestone-project-self-traing-loop-data",
    "phase1_all",
    split="phase1_all",
)

Consumers that need the historical split must join phase1/split.json to the Parquet rows by pair_id. They must not infer split membership from shard names or row order.

Record schema

Field	Type	Meaning
pair_id	string	Stable Phase 1 pair identifier
query	string	Generated question
chosen_prompt	string	Prompt associated with the historical preferred answer
rejected_prompt	string	Prompt associated with the historical rejected answer
chosen_answer	string	Historical preferred generated answer
rejected_answer	string	Historical rejected generated answer
chosen_f1	float64	Historical token-overlap F1 score
rejected_f1	float64	Historical token-overlap F1 score
signal_type	string	retrieval_helps or retrieval_hurts
data_source	string	Phase 1 provenance bucket
retrieved_documents	optional list of structs	Retrieved text, score, and index

Each retrieved_documents element has:

text: string
score: float64
index: int64

Historical limitations

1. The old Phase 1 split has no normalized-question crossing, but the later provenance audit found 78 exact normalized source passages crossing the historical train/validation boundary. The split is retained for exact Phase 1 reproduction only.
2. Phase 2 must construct a new near-deduplicated, grouped, leakage-safe split. This repository does not yet contain a frozen Phase 2 dataset.
3. chosen and rejected encode the historical Phase 1 automatic preference decision. Token-overlap F1/EM can disagree with semantic correctness and grounding; these fields must not be interpreted as human preference labels.
4. Generated questions and answers may be inaccurate, unsupported, or otherwise unsuitable for deployment.
5. The Parquet conversion found 1,847 records without retrieved_documents; the field is therefore nullable.

Integrity and provenance

The source JSON was not modified:

source path:
  outputs/phase1_preferences_all_unique/preference_pairs.json
source SHA-256:
  6e39bb5fa8bc6beaa499db4b58b04b093cfe00da850c6f0c4cbb6095f93abce9
source rows:
  111096

The publication conversion:

• preserves all 111,096 unique pair_id values;
• preserves the exact helps/hurts and source-bucket counts above;
• produces six Zstandard-compressed Parquet shards;
• verifies that the 99,986 historical train IDs and 11,110 historical validation IDs are disjoint and their union equals the published row IDs;
• uses MANIFEST.sha256 with paths relative to the repository root.

MANIFEST.sha256 deliberately does not hash itself. Verify the staged payload from the repository root with:

sha256sum --check MANIFEST.sha256

Excluded material

This release intentionally excludes:

• *.partial.jsonl, progress files, logs, caches, and temporary files;
• raw provider/API responses, request payload archives, admissions, private_mapping files, and terminal-failure records;
• model adapters, full weights, optimizers, checkpoints, and training caches;
• local credentials, shell configuration, and the project keys/ directory;
• all incomplete Phase 2 artifacts.

A scan of the selected Phase 1 source files found no common API-token, private-key, GitHub-token, Hugging Face-token, AWS-access-key, or Bearer-token signature. That limited signature scan is not a substitute for a full privacy/PII or copyright review.

Intended use

The intended uses are controlled research reproduction, provenance analysis, and development of the Phase 2 leakage-safe pipeline. The data are not approved for public redistribution, production model training, high-stakes decisions, or claims of human preference quality.