Project Adam (Phase 1: Synthetic Data Ingestion & Baseline Study)

Project Adam is an experimental framework designed to build an Autonomous Multi-Agent Social Ecosystem. The ultimate goal is to generate domain-specific synthetic social interactions, compress them into structural "Soul Documents," and train a Small Language Model (SLM) from scratch using sequential neural networks (LSTM) to mimic human behavioral biases, localized slang, and casual communication patterns.

This repository contains Phase 1 (V1), which focuses on engineering the automated synthetic data pipeline and establishing the initial raw baseline dataset.

---

V1 Dataset Architecture

The baseline dataset consists of 2,000 synthetic chat logs representing conversational interactions between two core personas (Marcel and Jessica). The pipeline successfully automated the data collection and structured the output into a unified, lightweight JSON format.

Data Preview (data_simulasi_tahun_1.json)

[
  {
    "index": 1,
    "year": "Tahun 1",
    "sender": "Jessica",
    "text": "ekhem udah lupa ada gw?"
  },
  {
    "index": 2,
    "year": "Tahun 1",
    "sender": "Marcel",
    "text": "hahahaha gw juga lupa sih, tapi gw ingat gw ada gw, so yeah, gw ada"
  }
]

---

Critical Evaluation & Engineering Insights (The V1 Imperfection)

As a robust Data Engineering and Machine Learning project, Phase 1 was strictly evaluated to analyze the behavioral limits of the local generative LLM framework.

Key Findings:

1. Context Looping & Mode Collapse: The extracted raw data exhibits a heavy pattern of repetition (e.g., hyper-fixation on localized loops like "capek bgt", "kesepian", and "shift sales").
2. Determinism vs. Entropy: Due to a low default temperature setting and lack of external state intervention during the simulation, the two agents fell into a conversational feedback loop, reducing the semantic variance of the dataset.
3. Implication for SLM Training: If this raw dataset is directly injected into the V3 LSTM architecture, the resulting Small Language Model will inherit this communication defect, producing highly predictable and repetitive responses.

---

The Blueprint Evolution (Next Steps)

The flaws discovered in this V1 baseline study directly inform the architectural design of Phase 2 (V2 Syn Theater) and Phase 3 (Mimicry).

[V1: Raw Ingestion (Current)] ──► [V2: Multi-Agent Syn Theater] ──► [V3: Custom SLM via LSTM]

Phase 2: Autonomous Theater Architecture

To eliminate context looping and introduce semantic entropy, Phase 2 will introduce:

• The Groq Director Layer: An external orchestration layer that dynamically injects real-time events and forces topic transitions every 10-15 tokens.
• Hyperparameter Tuning: Introducing frequency_penalty and scaling the temperature (0.85 - 1.0) to penalize repetitive phrasing and unlock creative vocabularies.
• The Infant Agent (Tabula Rasa): Injecting an unconditioned agent into the backchannels to absorb social jargon and behavioral biases through In-Context Social Learning.

Phase 3: Custom SLM Training (From Scratch)

• Extracting the refined chat logs of the Infant Agent to build a specialized localized Vocabulary Map (slang, typographies, emotional triggers).
• Building and training a custom Sequence-to-Sequence LSTM with Attention Gating from scratch in PyTorch to lock the sirkel's "soul" into an independent chatbot without relying on external LLM weights.

---

How to Inspect the Data

To run a quick diagnostics check on the harvested dataset, execute the following script:

import json

file_path = 'data_simulasi_tahun_1.json' # Adjust path if stored inside a subdirectory
with open(file_path, 'r', encoding='utf-8') as f:
    data = json.load(f)

print(f"Total Harvested Logs: {len(data)} entries.")
print(json.dumps(data[:5], indent=2, ensure_ascii=False))

---

Author: Adrian Marcello Budiman

Specialization: Computer Science / Machine Learning Infrastructure

Copyright © 2026 Adrian Marcello Budiman. All rights reserved under the MIT License.