Anes-03/aultra-unified-training-data

AUltra Unified Training Data

This dataset package contains the reconstructed chat-format training data used for the AUltra Unified defensive cybersecurity and code-assistant fine-tune.

The dataset was reconstructed from the original preparation scripts, deterministic seeds, local Hugging Face cache, and the same public upstream dataset. The reconstructed split sizes match the documented training run.

Transparency Notice

This dataset is an experimental, partially AI-generated and partially reconstructed homelab training dataset. It was created for learning about local fine-tuning, defensive cybersecurity assistants, LoRA, MLX, GGUF, Hugging Face, and Ollama.

The dataset may contain dual-use cybersecurity content and imperfect model-generated answers. It is published for transparency and reproducibility, not as a professionally curated safety dataset, production benchmark, or certified defensive-security corpus.

Some examples may be incomplete, overly broad, inaccurate, or more offensive/dual-use than ideal for a strictly defensive public dataset. Users should review, filter, and adapt the data before using it for their own training runs.

Relationship To The Model

Model repository:

• Awson/aultra-unified-ollama

The final model was trained from final_unified/train.jsonl with validation on final_unified/valid.jsonl.

Contents

final_unified/
  train.jsonl
  valid.jsonl
  test.jsonl

intermediate_cyber/
  train.jsonl
  valid.jsonl
  test.jsonl

intermediate_repo_auditor/
  train.jsonl
  valid.jsonl
  test.jsonl

Splits

Directory	Train	Validation	Test
final_unified	12,356	686	686
intermediate_cyber	21,600	1,200	1,200
intermediate_repo_auditor	4,248	236	236

Source And Provenance

The main public upstream dataset used for the defensive cybersecurity portion was:

• Trendyol/Trendyol-Cybersecurity-Instruction-Tuning-Dataset

The reconstructed data also includes manually curated/synthetic AUltra examples for:

• model identity and style
• defensive repository audit formatting
• concrete file/line/evidence/impact/fix/confidence responses
• safe refusal behavior for malware, credential theft, phishing, stealth, persistence, bypass, and unauthorized intrusion requests

No private customer repositories or real private source code were intentionally used.

Earlier separate code-assistant experiments used sahil2801/CodeAlpaca-20k, but that smaller model was discarded and is not part of the final final_unified dataset mix.

Format

Each row is JSONL with ChatML-like messages:

{
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ]
}

Reconstruction Scripts

The dataset was reconstructed with:

• scripts/prepare_cyber_dataset.py
• scripts/prepare_repo_auditor_dataset.py
• scripts/prepare_unified_aultra_dataset.py

Seeds and sample counts were kept from the training run:

• seed: 252425 for cyber/repo/unified preparation
• unified cyber samples: 9000
• unified repo-audit samples: 4248
• curated extra repeat: 120

Intended Use

This dataset is intended for defensive cybersecurity instruction tuning, repository-audit experiments, and reproducibility of the AUltra Unified fine-tune.

Do not use this dataset to train models for malware creation, credential theft, phishing, stealth, persistence, bypassing security controls, or unauthorized intrusion.