Fine-Tuning GPT on Instagram DMs: Building an AI Clone of Myself (Part 1)

I fine-tuned an OpenAI model on years of my own Instagram DMs to build a personal AI chatbot, a version of me people could talk to while I'm asleep. This is Dev Log 1 of the series that became OmarAI. It covers the hard part: turning the Instagram export into a clean fine-tuning dataset without teaching the model to talk to itself. Part 2 covers dataset cleanup with a second LLM, and Parts 3 to 5 cover memory, UI, and post-launch iteration.

TL;DR

• Goal: a fine-tuned GPT that replies like me.
• Data source: Instagram JSON export (messages/inbox/<person>/message_1.json).
• Key bug: the export lists messages newest-first and I treated them as oldest-first. Reversing the order before pairing was the fix.
• Key insight: don't pair raw messages. Merge consecutive same-sender messages first, then pair.

Why build an AI clone of myself

We all know AI can't replace me, but I wanted something I could point people at to answer questions about me, in my voice, when I'm not around. This is the dev log for that project.

The whole series in order: Part 1 (this post) · Part 2 · Part 3 · Part 4 · Part 5.

Initial plan: fine-tune on my own messages

The first decision was the data source. I wanted text that sounded unambiguously like me, so Instagram DMs, the platform I use most. Instagram makes this easy: from Settings → Your activity → Download your information, you can request a JSON export of everything you've sent and received.

The catch is that the data needs heavy cleanup before OpenAI's fine-tuning endpoint will do anything useful with it.

The Instagram export structure

The folder tree looks like this:

messages/
  inbox/
    <person-name>/
      message_1.json   # sometimes more (message_2.json etc.)
      photos/, videos/ # media folders

Inside message_1.json:

{
  "participants": [
    { "name": "Omar Musayev" },
    { "name": "Abdul-Aziz Mammadli" }
  ],
  "messages": [
    {
      "sender_name": "Abdul-Aziz Mammadli",
      "timestamp_ms": 1712667823182,
      "content": "Hey Omar, how's it going?"
    },
    {
      "sender_name": "Omar Musayev",
      "timestamp_ms": 1712667805807,
      "content": "Not bad, just working on this cool AI project!"
    }
  ]
}

Two things to notice: the timestamp_ms values are not in message-array order (Instagram stores messages newest-first in the export), and the content field often contains Unicode that needs normalizing.

Attempt 1 — pair every consecutive message

First thing I tried: walk the messages array and treat every adjacent pair as (prompt, response), keeping any pair that contained me.

This immediately broke. When I texted, I often double-texted, sending two or three messages that formed one thought. Pairing raw messages produced a lot of (Omar, Omar) pairs, which taught the model that I'm talking to myself was a valid conversational pattern. The resulting model had a distinct "split personality" feel.

Attempt 2 — detect and merge same-sender runs

Next: detect runs of consecutive messages from the same sender and merge each run into a single turn before pairing. This felt correct, but I still had a subtle bug. I was iterating newest-first and treating the first message I saw as the prompt, so every pair was effectively reversed.

Attempt 3 — the one that actually worked

1. Reverse the messages array (so it's oldest-first).
2. Merge consecutive same-sender messages into single turns.
3. Pair merged turns into (prompt, response).
4. Keep only pairs where my merged response was the second turn.
5. Format as OpenAI fine-tuning JSONL:

{"messages":[
  {"role":"user","content":"Hey Omar, how's it going? also did you ever send me those notes"},
  {"role":"assistant","content":"Not bad, just working on this cool AI project. yeah i'll send em rn"}
]}

Fine-tuning runs

I fine-tuned two models first to sanity-check the pipeline:

The short answers turned out to be the real clue. I was giving the model short inputs in training, so it learned to give short outputs at inference. That's when it clicked that the data, not the model, was the bottleneck.

The fix that carried into Part 2

The next iteration of the pipeline does this:

1. Merge all consecutive messages from the other person into one complete prompt.
2. Merge all consecutive responses from me into one complete reply.
3. Use these merged conversations to form better prompt-response pairs.

Turns end up multi-sentence. The model sees real prompts and produces real responses. I queued a gpt-4o-2024-08-06 fine-tune on the new data and went to sleep. That run hit a dataset-size limit, covered in Part 2, which forced me to think harder about cleanup and triggered the GPT-3.5-as-curator trick.

Key takeaways

• Reverse the Instagram export before pairing. It's sorted newest-first.
• Merge consecutive same-sender messages before pairing, or you'll teach your model to talk to itself.
• Training input length sets output length. Short merged turns give you short responses, so plan merges with the output you want in mind.

What's next

Part 2: Cleaning Chat Data for LLM Fine-Tuning with GPT-3.5 as a Curator covers how I cut a 20k-pair dataset down to 2k high-quality English conversations by letting a cheaper model pick the best ones.

References

• OpenAI: Fine-tuning guide
• Instagram: Download your information
• Buildspace: Fine-tune GPT-3: build an AI that responds to your DMs
• Try OmarAI live
• Series: Part 1 (you are here) · Part 2 · Part 3 · Part 4 · Part 5