README.md

metadata

license: apache-2.0
language:
  - en
tags:
  - autism-support
  - neurodiversity
pretty_name: Autism consultant
size_categories:
  - 10K<n<100K

---
license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- neurodiversity
- autism-support
- clinical-psychology
- sft
- chatml
size_categories:
- 10K<n<100K
---

Dataset Card for Adult Autism & Sensory Systems Consulting Dataset

This dataset is a high-density, highly scrubbed Supervised Fine-Tuning (SFT) corpus designed to train language models in adult autism advocacy, workplace environmental engineering, neurodiversity-affirming psychology, and tactical executive dysfunction adjustments.

The entire dataset is natively formatted in the ChatML schema, ensuring immediate compatibility with modern training frameworks like Unsloth, Axolotl, or alignment scripts.

---

Dataset Description

• Primary Domain: Adult Autism Clinical & Practical Consultation
• Format: JSON Lines (.jsonl) utilizing standard ChatML turn architectures
• Total Records: 16,000 high-density QA pairs (12,000 Expert Reference + 4,000 Algorithmic Synthesized)
• Core Objective: Eliminating conversation padding, repetitive pleasantries, and meta-textual document references to force direct, clinical, and authoritative structural protocols.

---

Dataset Structure & Schema

Each row within the JSONL file represents a complete standalone conversation sequence containing a hard-coded system persona instruction, a user problem statement, and an authoritative assistant response.

{
  "messages": [
    {"role": "system", "content": "You are a professional research model. Answer the user prompt strictly, exhaustively, and with academic depth."},
    {"role": "user", "content": "[Direct, scenario-driven or structural query regarding neurodivergent profiling or accommodation]"},
    {"role": "assistant", "content": "[Dense, zero-filler, multi-point tactical engineering plan or deep clinical explanation]"}
  ]
}

---

Data Curation & Pipeline Engineering

The high structural predictability of this dataset is the result of an intentional two-stage python pipeline designed to enforce linguistic diversity while aggressively cutting down on semantic redundancy.

1. Synthesized Augmentation Pipeline (QA_dataset_generator.py)

To complement the 12,000 primary expert reference entries, 4,000 scenario-grounded rows were algorithmically synthesized from structured markdown repositories using an instruction-tuned local instance. To prevent the model from generating monotonous queries, the synthesis script enforces a randomized assignment across six distinct linguistic archetypes for every generation turn:

• Situational Case Presentations: Frames a brief, practical scenario or real-world manifestation before appending a targeted inquiry.
• Myth-Correction Matrix: Frames prompts around popular misunderstandings, historical fallacies, or points of skepticism requiring fact-based corrections.
• Cross-Disciplinary Links: Demands direct structural comparisons or analytical contrasts between distinct entities, methodologies, or concepts.
• Practitioner Inquiries: Mimics natural inquiries originating from industry professionals, field researchers, or stakeholders noticing specific real-world patterns.
• Internal Mechanics Deep-Dives: Centers the prompt explicitly around underlying processes, direct functional experiences, or structural pathways.
• Advanced Operational Requests: Bypasses traditional question marks entirely, using direct, advanced operational commands (e.g., "Trace the chronological shift in...").

2. Vector Anomaly Extraction Pipeline (cleaner_ChatML.py)

To prevent the model from absorbing low-quality token sequences, synthetic data outputs are post-processed through an embedding-based isolation layer:

• Embedding Model: Text vectors are generated using a local instance of snowflake-arctic-embed2:568m.
• Style Centroid Computation: The script calculates a global average vector across valid rows to determine the baseline mathematical "style signature" of the domain text.
• Cosine Similarity Sanitization: Every row is evaluated via cosine similarity against the global style centroid. Any row yielding an alignment score below 0.34 is categorized as an outlier anomaly and automatically dropped from the final training split.

---

Intended Use Cases

• Supervised Fine-Tuning (SFT): Teaching compact base models (e.g., 7B, 8B, 14B architectures) to speak with dense clinical accuracy rather than generic conversational chatter.
• Retrieval-Augmented Generation (RAG): Serving as a high-quality ground-truth reference database for encoding clinical and physical accommodation strategies.
• Academic/Educational Research: Studying the behavioral interaction patterns of structural workplace adjustments.

---

Considerations & Disclaimers

Clinical Disclaimer: This dataset contains structural strategies, physical environment parameters, and cognitive optimization protocols. It is intended strictly for instructional tuning, technical enablement, and informational tooling. It is not a diagnostic instrument, medical advice database, or a replacement for licensed clinical healthcare professionals.