llmitate

Fine-tune a local model to text like one specific person, using your Instagram DM export. Everything stays on your machines.

How it works

• Base model, not an assistant. A base LLM (default Qwen2.5-7B) is fine-tuned on the chat as a plain transcript and simply continues it. There is no system prompt and no persona stats to parrot.
• Every message is its own line, so the model learns her bursts of short messages instead of one blob. Session headers (--- Wed 14 Jan 2026, 11:41 PM ---) give it time-of-day awareness.
• Loss only on her messages (and on the speaker label after her lines, which teaches when a burst ends).
• Emojis are preserved byte for byte. Dropped: "Liked a message/reel", reactions, attachment notices, calls, unsends, nickname/collection events. Shares and media become [reel], [photo]... for context only.
• Chronological validation split, so held-out data never leaks into training.
• transcript.py is the single definition of the format used by training, the app and the evaluator.

Layout

File	Purpose
parse_export.py	Export JSON -> data/processed/conversation.jsonl
build_dataset.py	-> train.jsonl, val.jsonl, val_sessions.jsonl, pii_review.jsonl
train.py	QLoRA on the GPU machine -> models/adapters/final
export.py	Merge + GGUF + Modelfile
generate.py	Ollama client returning a burst of messages
app.py	Streamlit chat + Turing test
evaluate.py	Real-vs-generated style stats on held-out sessions

Workflow

1. Data machine: put message_*.json in data/raw/, then

   python -m venv .venv && source .venv/bin/activate && pip install -r requirements.txt
   cp config.yaml.example config.yaml        # set target_sender / my_sender
   python parse_export.py && python build_dataset.py
   pytest tests

   Check the printed sample window and data/processed/pii_review.jsonl, then git push the code.
2. GPU machine (Windows, RTX 5070): git pull, copy data/raw/*.json and config.yaml over manually.

   python -m venv .venv; .venv\Scripts\activate
   pip install torch --index-url https://download.pytorch.org/whl/cu128
   pip install -r requirements-train.txt
   python parse_export.py; python build_dataset.py
   python train.py --max-steps 20      # smoke test
   python train.py                     # full run, best checkpoint by val loss
   python export.py                    # GGUF + Modelfile

   If Unsloth is troublesome on Windows, use WSL2 with the same steps.
3. Run it, either:
   • on the GPU machine: ollama create divyanshi -f Modelfile, serve with OLLAMA_HOST=0.0.0.0, and set ollama_url in the laptop's config.yaml to http://<windows-ip>:11434; or
   • on the laptop: copy models/gguf/divyanshi-q4_k_m.gguf and Modelfile, then ollama create divyanshi -f Modelfile (4 GB VRAM means partial CPU offload, a few tokens/s, fine for short messages).
4. streamlit run app.py for chat and the Turing test (accuracy near 50% = indistinguishable), python evaluate.py for style distributions.

Tuning

• Replies too samey: raise temperature (0.9) or lower min_p. Too chaotic: lower temperature.
• Wrong thread names: parse_export.py prints counts per sender and warns on mismatch.
• Out of memory: batch_size: 1, raise gradient_accumulation_steps, or lower max_seq_length and window_chars.
• Overfitting (val loss rising): fewer epochs; load_best_model_at_end already keeps the best checkpoint.