Soulkiller

Compress yourself into an AI model to be immortal, trained on lifetime personal data from Gmail, Calendar, Drive, YouTube, Chrome, Messenger, Instagram, WhatsApp, Telegram, X, LinkedIn, Reddit, notes and git history.

Download model GGUF · 4.7 GB · Hugging Face Source code michaljach/soulkiller
The purpose of Soulkiller was to digitize the minds of people, creating engrams—consciousness transferred into coding form.
Cyberpunk Wiki

How it works

  1. 01

    Export

    Download a lifetime of data from Google, Meta, WhatsApp, Telegram, X, LinkedIn and Reddit, plus notes and git history.

    data/raw/
  2. 02

    Ingest

    Parsers for 14 sources normalize every export into messages and documents, strip quoted replies, redact passwords, codes and card numbers, and drop newsletters and spam.

    messages.jsonl docs.jsonl
  3. 03

    Build

    Splits threads into conversation sessions and turns each into a chat example: your replies are the training targets, everyone else's messages are context. Holds out 3% of threads for validation.

    sft_train.jsonl
  4. 04

    Train

    QLoRA fine-tune of Qwen3 8B on a single RTX 4080 Super, then export for local inference.

    LoRA adapter GGUF
  5. ◆

    Soulkiller

    A model with your personality and, soon, your memory, running entirely on your own hardware.

Model specs

Base model
Qwen3 8B
Method
QLoRA, 4-bit base
LoRA
rank 32, alpha 32, all attention and MLP projections
Context
4,096 tokens
Training
2 epochs, lr 1e-4 cosine, AdamW 8-bit, bf16
Batch
2 × 8 gradient accumulation
Stack
Unsloth, TRL, Hugging Face Datasets
Hardware
RTX 4080 Super, 16 GB
Export
LoRA adapter, GGUF Q4_K_M

Benchmarks

Validation loss

Loss on held-out replies during training. Lower is better.

Data

Perplexity

How surprised each model is by real held-out messages. Lower is better.

Data

Blind test

How often friends spot the real person. 50% means they can't tell the two apart.

Data

Inference speed

Tokens per second as the conversation grows, GGUF Q4_K_M on the RTX 4080 Super.

Data

Memory

GPU memory used at each context length. Lower is better.

Data

Training data

Tokens written by the person, by source.

Data

Personality

A QLoRA fine-tune of Qwen3 8B on messages you wrote yourself: email replies, chats and posts. Training loss is applied only to your side of each conversation. Other people's messages appear only as the context you were replying to.

Memory

Next, a retrieval index over everything: mail, documents, calendar, notes and location history. The model then has real facts to draw on.

Data

  • Gmail, Calendar, Drive, YouTube, Chrome
  • Messenger, Instagram, WhatsApp, Telegram
  • X, LinkedIn, Reddit
  • Notes and git commit messages

Privacy

Everything stays local: parsing and training run on your own hardware, and a single 16 GB GPU is enough. Passwords, codes and card numbers are redacted at parse time.