GenAIDevTOProd/databricks-dolly15k-semantic-complexity

Databricks - Dolly 15k – Enriched Variant (Instruction-Tuned with Semantic and Complexity Augmentation)

Overview

This dataset is a semantically enriched and complexity-aware extension of the original Databricks Dolly 15k, purpose-built for evaluating and training instruction-following models. Each sample is augmented with additional signals to enable more nuanced filtering, curriculum learning, and benchmark development across diverse NLP tasks.

Dataset Format

Each sample includes the following fields:

instruction (str) – the prompt or task instruction

context (str) – optional background text (can be empty)

response (str) – the ideal output corresponding to the instruction

category (str) – the original category label (e.g., closed_qa, generation)

category_enriched (List[str]) – multi-class enriched taxonomy based on LLM-based relabeling

embedding (List[float]) – 384-dimensional semantic representation using a SentenceTransformer model

instruction_readability (float) – Flesch Reading Ease score of the instruction

response_readability (float) – Flesch Reading Ease score of the response

instruction_tokens (int) – number of tokens in the instruction

response_tokens (int) – number of tokens in the response

Enrichment Details

1. Semantic Embeddings (384-D)

Model: all-MiniLM-L6-v2 from SentenceTransformers

Purpose:

Enables vector search and similarity-based retrieval.

Facilitates clustering or curriculum grouping based on semantic distance.

Use Case: RAG pipelines, hybrid retriever-generator evaluation, semantic data deduplication.

2. Multi-Label Category Enrichment

Method: LLM-based enrichment of original category into multiple labels reflecting nuanced intent (e.g., closed_qa, classification, instruction_reformulation).

Purpose:

Allows for filtering and training multi-task models with fine-grained task type supervision.

Enables few-shot sampling or balanced evaluation subsets.

Use Case: Model generalization studies, task disambiguation training, LLM taxonomy alignment.

3. Readability Scores

Metric: Flesch Reading Ease

Range: Typically from -10 (very complex/short text) to 100+ (very easy to read)

Interpretation:

Higher is simpler: A score above 60 indicates easy-to-read content.

Negative values: Usually from one-word answers or very short instructions.

Purpose:

Measures linguistic complexity for curriculum learning.

Enables filtering of prompts based on difficulty level.

4. Token Lengths (Instruction/Response)

Method: tiktoken tokenizer for gpt-3.5-turbo vocabulary

Purpose:

Supports token-level curriculum filtering.

Enables outlier detection for unusually long or short samples.

Use Case: Model length conditioning, latency profiling, instruction tuning length analysis.

Research Use Cases

Curriculum Learning: Use readability and token length to gradually train models from simple to complex examples.

Semantic Similarity Evaluation: Leverage embeddings for nearest-neighbor search, duplicate detection, or hybrid retriever training.

Task-Type Robustness: Train and evaluate models across enriched multi-label categories to assess generalization across QA, classification, and generative tasks.

Prompt Engineering Validation: Analyze impact of prompt complexity (via readability/tokens) on response quality.

Citation

@online{DatabricksBlog2023DollyV2, author = {Mike Conover and Matt Hayes and Ankit Mathur and Jianwei Xie and Jun Wan and Sam Shah and Ali Ghodsi and Patrick Wendell and Matei Zaharia and Reynold Xin}, title = {Free Dolly: Introducing the World's First Truly Open Instruction-Tuned LLM}, year = {2023}, url = {https://www.databricks.com/blog/2023/04/12/dolly-first-open-commercially-viable-instruction-tuned-llm}, urldate = {2023-06-30} }

Wikipedia (various pages) - https://www.wikipedia.org/ Copyright © Wikipedia editors and contributors.

License

Same as original Dolly 15k: Creative Commons Attribution-ShareAlike 3.0 (CC BY-SA 3.0)