LLM-Imitate

Fine-tune a local open-weight LLM to imitate someone's Instagram DM texting style — 100% offline, no cloud APIs.

---

What it does

1. Parses your Instagram DM export into a structured conversation timeline.
2. Profiles the target sender's writing style (emoji habits, slang, sentence length, punctuation patterns).
3. Builds a fine-tuning dataset of (context → reply) pairs in Llama 3 chat format.
4. Fine-tunes Llama 3.1 8B Instruct with QLoRA (via Unsloth) on your local GPU.
5. Exports the merged model to GGUF (Q4_K_M) and registers it in Ollama.
6. Serves an interactive chat UI with Streamlit.

Everything runs locally after the initial model download. Your data never leaves your machine.

---

Requirements

Requirement	Notes
Python 3.10+
NVIDIA GPU	8 GB+ VRAM recommended (tested on 12 GB)
CUDA 12.0+	CUDA 13 for Blackwell/RTX 50xx
Ollama	For serving the final model

---

Quick start

1. Clone and install

git clone https://github.com/YOUR_USERNAME/LLM-Imitate.git
cd LLM-Imitate

python -m venv .venv
.venv\Scripts\activate        # Windows
# source .venv/bin/activate   # Linux / macOS

pip install -r requirements.txt

# Install PyTorch with CUDA support (adjust cu130 to your CUDA version)
pip install torch --index-url https://download.pytorch.org/whl/cu130
pip install unsloth

2. Export your Instagram DMs

In the Instagram app: Settings → Account → Your activity → Download your information. Select JSON format, then unzip and copy the thread folder into:

data/raw/inbox/<thread-name>/
  message_1.json
  message_2.json
  ...

3. Configure

cp config.yaml.example config.yaml

Edit config.yaml:

Key	Description
thread_path	Path to the thread folder inside data/raw/inbox/
target_sender	Exact sender name to imitate (as it appears in the JSON export)
my_sender	Your name in the thread
max_train_samples	Set to a small number (e.g. 200) for a quick smoke test; null for full training

4. Run the pipeline

# Step 1 — Parse the raw export
python parse_export.py

# Step 2 — Build a style profile for the target sender
python build_persona_profile.py

# Step 3 — Build the fine-tuning dataset (review pii_review.jsonl first!)
python build_dataset.py

# Step 4 — Fine-tune (smoke test: set max_train_samples: 200 in config.yaml first)
python train.py

# Step 5 — Export to GGUF and register with Ollama
python convert_to_gguf.py
ollama create persona-imitate -f Modelfile

# Step 6 — Launch the chat UI
streamlit run app.py

# Step 7 (optional) — Evaluate style fidelity
python evaluate.py

---

Pipeline overview

Step	Script	Output
1	parse_export.py	data/processed/conversation.jsonl
2	build_persona_profile.py	data/processed/persona_profile.json
3	build_dataset.py	train.jsonl, val.jsonl, pii_review.jsonl
4	train.py	LoRA adapter → models/adapters/final/
5	convert_to_gguf.py	Quantized GGUF → models/gguf/
6	Ollama + Modelfile	Local inference model persona-imitate
7	app.py	Streamlit chat UI at http://localhost:8501
8	evaluate.py	Style-fidelity report

---

Project structure

LLM-Imitate/
├── config.yaml.example     # Copy to config.yaml and fill in your details
├── data/
│   ├── raw/                # Instagram export — gitignored
│   └── processed/          # Generated artifacts — gitignored
├── models/
│   ├── adapters/           # LoRA checkpoints — gitignored
│   └── gguf/               # Merged + quantized weights — gitignored
├── parse_export.py
├── build_persona_profile.py
├── build_dataset.py
├── train.py
├── convert_to_gguf.py
├── Modelfile
├── app.py
├── evaluate.py
└── utils.py

---

Training tips

• Smoke test first: Set max_train_samples: 200 and epochs: 1 in config.yaml to verify the pipeline end-to-end in ~5 minutes before committing to a full run.
• Full run: Set max_train_samples: null and epochs: 3. Expect ~1–3 hours depending on your dataset size and GPU.
• PII: Always review data/processed/pii_review.jsonl before training — it flags messages containing phone numbers, emails, and similar patterns.
• Overfitting: Keep epochs at 2–3 and watch the validation loss. With 30k+ source messages you should have plenty of data.
• GPU memory: If you see Unsloth: Will smartly offload gradients, reduce max_seq_length to 1024 and set packing: false in config.yaml.

---

Configuration reference

See config.yaml.example for all available options with inline comments.

---