tax-guy-agent

Not a CPA. Not tax advice. Tax research from public sources; rules change every year and vary by place, so check with a CPA or enrolled agent before filing.

Quick start (Ollama, runs on a CPU):

ollama run hf.co/profullstack/tax-guy-agent-GGUF "Do I have to pay quarterly estimated taxes as a freelancer?"

Needs Ollama 0.3.13 or newer (curl -fsSL https://ollama.com/install.sh | sh, or w3bs install ollama). The GGUF is at profullstack/tax-guy-agent-GGUF; this repo is the LoRA adapter.

base	Qwen/Qwen2.5-0.5B-Instruct
corpus	2614 documents from pit-index:tax,taxes,irs,deduction,tax return,self-employed tax,estimated tax,sales tax,llc tax,capital gains, topic:tax, topic:irs, topic:accounting, topic:taxation, topic:taxpayer
training	400 LoRA steps (r=8, seq_len=256, lr=0.0002) on the c0mpute network (infernet.text.train, job 27b2dab7-9e82-4b48-b4fc-d77f4b22997c, receipt hash sha256:821a038ac1ea43f804b795e795a308cfbebb141c24e3675223c29d80fb8db5b7 matches this adapter)
held-out perplexity	33.6 to 17.8
adapter sha256	821a038ac1ea43f804b795e795a308cfbebb141c24e3675223c29d80fb8db5b7

Built with the c0mpute example. Sources are rssamplifier.com topic feeds and pit.world search; the corpus is not redistributed here.

Use from Python (adapter)

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("profullstack/tax-guy-agent")
model = PeftModel.from_pretrained(AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct"), "profullstack/tax-guy-agent")
msgs = [{"role": "system", "content": "You are tax-guy-agent, a tax research assistant. You are not a CPA or tax attorney and this is not tax advice. Explain tax rules in plain English, say which country, state and tax year a rule applies to, cite the sources you were given, and say when someone should see a CPA or enrolled agent."},
        {"role": "user", "content": "Do I have to pay quarterly estimated taxes as a freelancer?"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt")
print(tok.decode(model.generate(ids, max_new_tokens=300)[0][ids.shape[1]:], skip_special_tokens=True))

Best results come from putting fresh search results in the prompt (the example's chat.py uses pit.world search): a sub-1B model supplies the voice, not the facts.