custom-ammar-agent

Custom Ammar Agent Dataset

This dataset is designed for fine-tuning language models (e.g., Qwen2.5-Coder-7B-Instruct) to create an AI agent that understands the Ammar Jewels project (a Nuxt 4 + Drizzle/Turso monorepo) and can assist with coding, architecture, and domain-specific questions.

Dataset Composition

The dataset is a mixture of three sources, formatted as ShareGPT-style JSONL (each line is a JSON object with a "conversations" array containing user and assistant turns).

1. Chat (chat.jsonl)
   Extracted from real conversation logs (7 sources: 4 .md exports and 3 .json attachments).
   Contains technical discussions, architecture decisions, and debugging sessions in Bahasa Malaysia and English.
   Cleaned to remove system noise, tool calls, and non-essential content.

2. Docs (docs.jsonl)
   Auto-generated question-answer pairs from project documentation:

   • MASTERPLAN.md
   • AGENTS.md
   • docs/AUDIT-CORE-VERTIKAL.md
   • docs/BIO-SITE-BUSINESS-PLAN.md
   • docs/FRAPPE-BLUEPRINT.md
   • docs/GUARDRAILS.md
   • docs/PERLEMBAGAAN.md

   Questions are auto-templated (e.g., "Terangkan tentang: ...", "Apakah Fasa X?") and answers are the original text chunks.

3. Code (code.jsonl)
   Extracted from write_file and patch tool calls in the conversation logs.
   Each example consists of a natural language request (inferred from context or generated) and the actual code that was written or patched.
   Covers: API endpoints, composables, utilities, components, schema modifications, and more.

Splits

Note: After deduplication, 759 unique examples remain from a gross of 762.

Format Example

{"conversations":[{"role":"user","content":"Buat endpoint POST untuk stok/pembelian?"},{"role":"assistant","content":"import { eq } from 'drizzle-orm'\\nimport { stocks, purchases } from '@ammar/database'\\n\\ndefineEventHandler(async (event) => {\n  const body = await readBody(event)\n  // ... validation and DB insert\n})"},"source":"ui1","tool":"write_file","path":"apps/portal/server/api/stock/purchases.post.ts","lines":42}

Usage

The dataset is ready for supervised fine-tuning (SFT) or LoRA/QLoRA training.
Example with �� 🤗 Transformers and �� 🤗 PEER:

from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments, Trainer

dataset = load_dataset("syaher/custom-ammar-agent")
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-Coder-7B-Instruct")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-Coder-7B-Instruct")

# ... tokenize, set up Trainer, etc.

Generation Notes

• All personally identifiable information (PII) such as real phone numbers, emails, passwords, and API keys have been removed.
• The dataset is primarily in Bahasa Malaysia (reflecting the project's language), with some English technical terms.
• Created from the nuxt-template repository (Ammar Jewels monorepo) as of August 2026.
• For reproducibility, see the extraction scripts in scripts/dataset/ of the source repository.

License

This dataset is released under the CC-BY-SA 4.0 license. Please attribute appropriately if used.

---

Generated automatically via dataset creation pipeline. For questions or issues, contact the dataset curator.