laatey's picture
Update README.md
584fbca verified
|
Raw History Blame Contribute Delete
2.67 kB
metadata
license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - distillation
  - open-weights
  - reasoning
  - CoT
  - instruction-tuning
  - synthetic-data
size_categories:
  - 100K<n<1M

definatelycustomdataset (Unified Open-Model Distillations & Reasoning Traces)

laatey/definatelycustomdataset is a curated, high-density collection of multi-turn conversations, reasoning chains, and instruction traces distilled directly from top-tier open-weight foundation models (e.g., Qwen 2.5, DeepSeek, and Llama 3).

Specifically engineered for supervised fine-tuning (SFT) of small-to-medium parameter models (100M to 7B), this dataset emphasizes clear step-by-step reasoning, instruction following, and clean output formatting without proprietary licensing constraints.


💡 Why This Dataset?

Many open distillation sets suffer from noisy outputs, "As an AI language model" boilerplate, broken schemas, or conflicting data formats. This dataset resolves those bottlenecks:

  • 100% Open Weights Distilled: Created without dependency on closed proprietary APIs, ensuring clean licensing and auditability.
  • Dense Chain-of-Thought (CoT): Emphasizes problem-solving, code explanation, mathematics, and structured logic.
  • Ready for Low-Resource SFT: Sharded and typed specifically to prevent out-of-memory issues or loader crashes when training small models (e.g., Nanowhale, SmolLM, Qwen-0.5B/1.5B/3B, Llama-3.2).

🛠️ Data Engineering & Pipeline

Each sample in this dataset has undergone rigorous normalization:

  1. Strict Message Schema: Converted raw text, prompt/response pairs, and ShareGPT format into the standard Hugging Face / OpenAI messages format ([{"role": "user", ...}, {"role": "assistant", ...}]).
  2. PyArrow / Parquet Validation: Enforces rigid List[Dict[str, str]] typing, preventing the common Couldn't cast array errors across modern loaders (TRL, Axolotl, Unsloth, LLaMA-Factory).
  3. Quality & Heuristic Filtering:
    • Removed empty, repetitive, or single-word turns.
    • Filtered out generic refusal tropes and synthetic artifacts.
    • Deduplicated identical prompt-response pairs.

📊 Dataset Structure

Each row is structured as a standard conversation:

{
  "messages": [
    {
      "role": "user",
      "content": "Explain how a distributed cache prevents cache stampede, with pseudo-code."
    },
    {
      "role": "assistant",
      "content": "A cache stampede (or thundering herd problem) occurs when high-concurrency requests hit an expired key simultaneously...\n\n### Mitigation Strategy: Mutex / Distributed Lock\n```python\n..."
    }
  ]
}