🍵 Tea

Tea is a full fine-tune of microsoft/phi-4, built for question answering and long, sustained assistant-style conversations. It is multilingual, with fine-tuning focused on strong, natural Farsi (Persian) conversational ability, while retaining Phi-4's general English and multilingual competence.

Model Details

Base model: microsoft/phi-4 (14B parameters)
Fine-tuning method: Full fine-tune — all parameters updated, no LoRA/PEFT adapters
Fine-tuned by: artindnr
License: MIT
Languages: Farsi (primary conversational focus), English, and general multilingual support
Model type: Causal decoder-only chat/assistant language model

What's New

tea takes Phi-4's strong base reasoning and language capabilities and tunes them specifically for:

Question answering — direct, accurate answers grounded in the conversation context
Long assistant conversations — maintaining coherence, tone, and context over extended multi-turn sessions rather than short single-shot exchanges
Farsi fluency — natural, idiomatic Persian conversation and assistance, alongside solid English and multilingual performance

Unlike adapter-based fine-tunes, every weight in the model was updated during training, which the author has found gives more consistent behavior for long-conversation use cases than LoRA-based approaches.

How to Use

Tea uses the standard chat template shipped with the base model, so it works out of the box with 🤗 Transformers.

Generation

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = "artindnr/tea"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

USER_PROMPT = "تو کی هستی و اسمت چیه؟"

messages = [
    {"role": "user", "content": USER_PROMPT},
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=1024,
    temperature=0.7,
    do_sample=True,
)

print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

For multi-turn conversations, simply keep appending {"role": "user", ...} / {"role": "assistant", ...} turns to the messages list before re-applying the chat template — tea is tuned to stay coherent as this history grows.

Intended Use

tea is intended for:

Farsi-first conversational assistants that also need to handle English/multilingual input
Question-answering applications requiring direct, grounded answers
Long-running, multi-turn assistant deployments (support bots, tutoring, general-purpose chat) where conversational memory and coherence over many turns matters
Research comparing full fine-tunes vs. adapter-based (LoRA) fine-tunes on the same base model

Limitations

As a full fine-tune, tea's Farsi-focused training may shift some of Phi-4's original English-centric behaviors; for English-only, general-purpose use cases the base microsoft/phi-4 model may still be preferable.
tea inherits the general capabilities and limitations of the phi-4 base model, including the possibility of hallucinated facts, especially over very long contexts.
No formal safety fine-tuning beyond what is inherited from the base model has been applied; use appropriate safeguards in production settings.

License

This model is released under the MIT License, consistent with the base microsoft/phi-4 model.

Citation

If you use tea in your work, please cite:

@misc{tea,
  title  = {tea: A Farsi-Focused, Full Fine-tune of Phi-4 for QA and Long-form Assistance},
  author = {artindnr},
  year   = {2026},
  url    = {https://huggingface.co/artindnr/tea}
}