Project 02 · Model layer · Shipped

Polar Orca

Formerly IRA Dolphine · Built by Rajiv Buda

I'm Polar Orca, built by Rajiv Buda to turn knowledge into opportunity, for everyone, everywhere.

A LoRA fine-tune of Qwen2.5-3B for short, practical advice. Trained on a free Colab T4, published as a 120 MB adapter, and written up with the limit in plain sight: one held-out prompt, not a benchmark.

Hugging Face model Model card on GitHub
Qwen2.5-3B-InstructBase model
120 MBLoRA adapter
<1 minTrain time, T4
$0Compute cost
<1%Weights trained

What Polar Orca stands for.

🌍Opportunity for everyone

Useful knowledge should not depend on where you were born or what you can pay.

🌉Bridges between people

Connect people to ideas, to work, and to each other, across languages and borders.

🤝Honest and humble

Say what was tested and what was not. No claims the evidence cannot carry.

📚Always learning

Each version is a step. Retrain, re-check, and publish what changed.

🐋Built by one person, for many

Built by Rajiv Buda on free tools, so anyone can reproduce it.

What I actually ran.

01Data

A small instruction set, seeded and then reviewed by hand. Not a scraped corpus.

02Train

Unsloth LoRA on 4-bit Qwen2.5-3B-Instruct (unsloth/Qwen2.5-3B-Instruct-bnb-4bit). Free T4, about 15 GB VRAM.

03Check

One prompt kept out of training, to see if the model only recited the set.

04Publish

Adapters pushed to the Hugging Face Hub. The URL is the artifact.

05Reload

The same model loads again from the Hub with no local checkpoint.

Decisions, and why.

Why Qwen2.5-3B

Small enough to quantize to 4-bit and train on a free T4, and still coherent on short strategic answers.

Why LoRA

The base stays frozen. The trainable piece is a low-rank adapter, about 120 MB instead of a full ~2 GB model.

Why Unsloth

Its kernels cut memory and time enough that the run fits the free Colab tier instead of a rented GPU.

Why 4-bit

NF4 shrinks the footprint enough to load and train on 15 GB. That is a hardware choice, not a quality claim.

The check I can show.

Held-out prompt · reloaded from the Hub · max_new_tokens=200
PromptGive strategic advice for a small fitness studio facing competition from a big gym chain.
Polar OrcaUnderstand your local market deeply and tailor your services to meet specific needs. Offer personalized coaching or unique classes that can't be replicated. Focus on building community relationships that a big chain can't match…

This prompt was not in the training set. The answer is coherent and on-domain. That is a smoke test of generalization, not a scored evaluation.

What this does not prove There is no held-out test set, no win-rate against the base model, and no claim that Polar Orca outperforms Qwen2.5-3B in general. A one-minute LoRA on a small instruction set can pick up a voice. It can also overfit. I am publishing the workflow because I ran it, not because a single paragraph is evidence of a better model.

Load it from the Hub.

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="maxrajiv-buda/polar-orca",
    max_seq_length=2048,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

messages = [{"role": "user", "content": "Give strategic advice for a small fitness studio facing competition from a big gym chain."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to("cuda")
out = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(out[0], skip_special_tokens=True))