Mungus451/verified_wiki_historian_the_beatles_anthology_dataset_active

Verified-Wiki-Historian: The Beatles Anthology

Verified-Wiki-Historian (The Beatles Anthology) is a refined, citation-grounded instruction dataset for Beatles-specific historical question answering, summarization, and supervised fine-tuning.

This dataset is a cleaned and rebuilt refinement of:

Mungus451/verified_wiki_historian_the_beatles_anthology_dataset_active

The current release contains 4,000 instruction records focused on Beatles history, recording sessions, release history, personnel credits, chart performance, film and media context, biographical context, business/legal history, and critical reception.

The dataset is designed for training or evaluating models that need to answer music-history questions with concise, source-aware responses rather than loose narrative recall.

---

Dataset Summary

Each record contains:

{
  "instruction": "...",
  "context": "...",
  "response": "...",
  "sources": ["..."],
  "topic": "...",
  "task_type": "...",
  "citation_marker_map": {
    "S1": "...",
    "S2": "..."
  }
}

The response field includes inline citation markers such as [S1], [S2], and [S3]. These markers correspond directly to entries in citation_marker_map, allowing the model or evaluator to trace claims back to the source list provided with the record.

Final validation of the refined release showed:

• 4,000 records
• 4,000 valid JSON object records
• 0 invalid JSON lines
• 0 non-object records
• 1 consistent schema
• 3,975 clean records
• 25 flagged records, all from the audit’s broad wiki_edit_artifact detector
• 0 citation-map mismatches observed in the final validation pass

The final audit showed matching citation-marker and citation-map distributions across the dataset, meaning inline markers and map entries are aligned record-by-record.

---

Relationship to the Original Dataset

This release is a refinement of the earlier active Beatles anthology dataset. The original dataset contained a larger set of records and a broader extraction footprint. This version prioritizes:

• cleaner JSONL structure
• consistent schema
• improved instruction phrasing
• rebuilt responses
• source-aware inline citation markers
• normalized citation maps
• removal of pipeline/debug fields
• reduced scrape artifacts
• stronger downstream training readiness

The goal of the refinement was not simply to increase record count. The goal was to produce a cleaner, more reliable dataset that can be used directly for supervised fine-tuning, LoRA/QLoRA experiments, and grounded QA evaluation.

---

Refinement Process

The dataset was processed through a multi-stage curation pipeline. The major stages included:

1. JSONL and Schema Normalization

Records were validated as JSON objects and normalized into a stable schema:

instruction
context
response
sources
topic
task_type
citation_marker_map

Pipeline metadata, temporary audit fields, and intermediate generation fields were removed from the final export.

2. Topic and Task Classification

Records were assigned a topic and task_type to make the dataset easier to filter, train on, and evaluate. The final task distribution includes:

Task Type	Records
personnel_credits	1,519
release_history	804
film_media_context	290
factual_summary	289
recording_history	255
chart_performance	243
biographical_context	189
composition_analysis	175
critical_reception	159
business_legal	77

This gives the dataset more variety than a simple generic summary corpus. It includes technical credits, recording history, release metadata, chart context, biographical summaries, criticism, and business/legal material.

3. Instruction Rewriting

The instructions were rewritten into clearer task-specific prompts. Instead of relying on repetitive generic prompts, the dataset now asks for specific outputs such as:

• summarize release history
• list known personnel and technical credits
• summarize recording history
• summarize chart performance
• summarize film, television, or media context
• explain composition and songwriting background
• summarize critical reception
• summarize Beatles-related business/legal context

This makes the dataset useful for training task-following behavior inside a specific historical domain.

4. Response Rebuilding

Responses were rebuilt to be concise, context-bound, and suitable for instruction tuning. The revised responses avoid unnecessary formatting noise and aim to answer only what is supported by the provided context.

When the context does not provide the requested information, the response generally states that limitation instead of inventing details. This is useful for training models to refuse unsupported specificity in a narrow historical domain.

5. Inline Citation Marker Injection

Citation markers were added directly into the response field. Each marker corresponds to a source entry in citation_marker_map.

Example structure:

"The song was recorded at EMI Studios in London. [S1]"

with:

"citation_marker_map": {
  "S1": "source text or source identifier"
}

This makes the dataset useful for:

• citation-aware SFT
• source-grounded response training
• grounded QA evaluation
• RAG-style answer comparison
• models that need to learn source attribution style

6. Citation-Map Cleanup

A major cleanup pass removed unused citation-map entries and ensured the inline markers and citation maps matched. In the final dataset, the citation-marker distribution and citation-map distribution match exactly:

Marker Count	Records
1	1,171
2	877
3	1,774
4	178

This matters because many citation datasets contain stale or unused source-map entries. This release avoids that problem.

7. Final Micro-Polish

The final passes fixed mechanical citation artifacts, including:

• unused citation-map entries
• citation markers split into title text
• escaped apostrophe artifacts
• malformed possessive title formatting
• remaining response-level formatting residue

The last cleanup pass changed only the response field, fixing 89 escaped apostrophes across 87 records, with no schema changes and no validation flags.

---

Dataset Composition

The dataset covers Beatles-related material across several recurring domains:

Category	Description
Recording History	Studio sessions, production context, engineers, overdubs, mixes
Personnel Credits	Performers, instruments, producers, engineers, technical roles
Release History	Albums, singles, B-sides, reissues, compilations, formats
Chart Performance	Chart positions, certifications, commercial reception
Composition Analysis	Songwriting, authorship, lyrical context, musical structure
Film and Media Context	Films, documentaries, broadcasts, music videos, television appearances
Biographical Context	Beatles members, collaborators, managers, producers, related figures
Critical Reception	Contemporary and retrospective criticism
Business and Legal Context	Apple Corps, publishing, management, disputes, releases

Top recurring topics include Paul McCartney, George Harrison, Ringo Starr, John Lennon, George Martin, The Beatles, Revolver, Rubber Soul, Abbey Road, Sgt. Pepper’s Lonely Hearts Club Band, The White Album, Beatlemania, Magical Mystery Tour, Strawberry Fields Forever, Hey Jude, and related Beatles works.

---

Intended Uses

This dataset is suitable for:

Supervised Fine-Tuning

Use it to fine-tune models on concise, grounded music-history answers.

Recommended formats:

• Alpaca-style instruction tuning
• ChatML conversion
• ShareGPT-style conversion
• LoRA / QLoRA training
• domain-adaptation SFT

Citation-Aware Answering

Because responses include inline [S#] markers and a citation map, the dataset can help train models to generate answers with source-style attribution.

Grounded QA Evaluation

The dataset can be used to evaluate whether a model answers from supplied context rather than relying only on parametric memory.

RAG Comparison

The dataset is useful for testing whether a model fine-tuned on grounded answers behaves similarly to a retrieval-augmented model when given compact context.

Music-History Assistants

This dataset is especially useful for models intended to answer questions about:

• Beatles songs
• album history
• recording sessions
• personnel credits
• release history
• chart performance
• Beatles-related films and media
• George Martin and production history
• Lennon-McCartney authorship questions

---

Recommended Training Format

A simple SFT prompt format:

### Instruction:
{instruction}

### Context:
{context}

### Response:
{response}

For chat-tuned models, the instruction and context fields can be combined into the user message, with response as the assistant message.

Example:

{
  "messages": [
    {
      "role": "user",
      "content": "Instruction: Summarize the recording history of \"A Day in the Life\".\n\nContext: ..."
    },
    {
      "role": "assistant",
      "content": "..."
    }
  ]
}

Recommended training settings depend on model size, but this dataset is well-suited for:

• 7B–14B models
• LoRA or QLoRA
• 2–4 epochs
• 2k–4k context windows
• low to moderate learning rates
• validation split for overfitting checks

---

Strengths

Clean Final Structure

The final release has a single consistent schema across all records and no invalid JSON lines.

Domain-Specific Depth

The dataset is narrow enough to teach a model Beatles-specific historical patterns, but broad enough to cover songs, albums, people, sessions, releases, charts, films, criticism, and business context.

Source-Aware Responses

Inline citation markers and citation maps make the dataset more useful than plain instruction/response pairs.

Reduced Hallucination Incentive

Many records explicitly stay within the provided context. When the context does not contain the requested details, the answer generally says so. This is useful behavior for models trained to avoid unsupported claims.

Practical Training Size

At 4,000 records, the dataset is small enough for fast LoRA experiments but large enough to provide meaningful domain adaptation.

---

Known Limitations

This dataset is not a complete Beatles encyclopedia. It is an instruction-tuning dataset derived from curated Beatles-related source material.

Known limitations:

• Some topics are more heavily represented than others.
• personnel_credits is the largest task category.
• Some responses are intentionally conservative when context is incomplete.
• A small number of audit flags remain from broad detection of the word “edit”; many are normal uses such as “final edit,” “composite edit,” or “edited version.”
• Citation markers are structurally aligned, but a few may still appear in stylistically imperfect positions inside titles or near punctuation.
• The dataset should not be treated as a replacement for primary-source research.

---

License

This dataset is released under:

Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0)

Because the dataset is derived from Wikipedia-style source material, downstream users should preserve attribution and comply with share-alike requirements.

---

Citation

If you use this dataset, cite the Hugging Face repository and credit:

Mungus451. Verified-Wiki-Historian: The Beatles Anthology. Hugging Face Datasets.

---

Dataset Credits

Data Architect: Mungus451 Project: Verified-Wiki-Historian Dataset: Verified-Wiki-Historian: The Beatles Anthology Base Dataset: Mungus451/verified_wiki_historian_the_beatles_anthology_dataset_active License: CC BY-SA 4.0 Domain: Beatles history, music history, citation-grounded QA, instruction tuning