khursanirevo/teach_aws

teach_aws — AWS Q&A in Bahasa Melayu (paraphrase-augmented)

Instruction-tuning data for answering AWS questions in Bahasa Melayu. Each row is a (question, answer) chat pair ready for SFT (TRL/axolotl-compatible messages format).

Built from PixelSpaceAI/aws-malay-qa (Apache-2.0):

• —Answers are verbatim from the source dataset — nothing was rewritten.
• —Rows are paraphrases only. Questions were paraphrased with a large language model in two passes:
• —para_v1 — neutral paraphrases of each source question.
• —para_v2_A … para_v2_D — typed paraphrases per source question covering syntactic rewording, consequence/inverse framing, terse informal register, and practical scenario framing.
• —The original Q/A pairs are not included — they are withheld as our private held-out evaluation set for on-going work, and are available from the source dataset anyway.

Row format

{"messages": [{"role": "user", "content": "..."},
              {"role": "assistant", "content": "..."}],
 "service": "sagemaker",
 "surface": "para_v2_B",
 "src_idx": 123,
 "src": "PixelSpaceAI/aws-malay-qa"}

service is the AWS service area as tagged in the source dataset (general, sagemaker, vpc, ...). surface identifies which variant of the question this row is.

Generation scope and controls

Every source question in the dataset was paraphrased; surfaces too close to another question were deduplicated by similarity. The original Q/A pairs are withheld as a private evaluation set and are available from the source dataset.

Why the methodology is trusted

Answer fidelity. Every assistant response is copied verbatim from the source dataset. Paraphrasing applies to questions only — answers were never generated, rewritten, or summarized, so the dataset cannot inject hallucinated technical content.

Paraphrase quality controls. Each paraphrase was required to preserve all technical terms (service names, feature names, formats) and the same information request. Two automatic filters applied to every generated surface: rejection of paraphrases too close to their source question (token-set Jaccard >= 0.75), and near-duplicate removal across questions.

No-leakage training methodology. This published pool covers every source question, including paraphrases of questions related to our held-out evaluation set. For our own fine-tuning we therefore use a filtered subset of the same generation output: only paraphrases of train-split questions, after screening every candidate surface — original or paraphrase — against the private held-out evaluation set with a token-set Jaccard firewall at 0.7 (any surface at or above the threshold is dropped). An independent verification step re-checks the final training file after every build; no training surface is a held-out question, a paraphrase of one, or near-duplicate of one (maximum retained similarity to date: 0.667, below threshold).

Generalization evidence. Fine-tuning on the decontaminated training subset improved accuracy on held-out questions over the base model, indicating the paraphrase augmentation teaches surface-invariance rather than memorization of evaluation items.

Statistics

Rows per AWS service (service field):

Attribution

• —Base questions and all answers: PixelSpaceAI/aws-malay-qa, Apache-2.0.
• —Paraphrases machine-generated from the source questions; answers unchanged.

Intended use / limitations

Malay-language AWS Q&A assistants and instruction tuning. Answers reflect the AWS FAQ content the source dataset was built from and may lag current AWS behavior; verify anything safety- or cost-critical against official AWS documentation.