tara1.2-quest

Tara 1.2 Quest

Tara 1.2 Quest is a tiny experimental question-generation model. It is the successor experiment to aungkomyint/tara1.1, with the task narrowed to producing JSON question lists for a user topic, keyword, sentence, or short request.

The model is designed to answer in this shape:

{"questions":["...","...","..."]}

It is not a general chat assistant. It is a small research model for structured question generation.

Model Details

• Model name: tara1.2-quest
• Internal checkpoint: tara-1.2-quest-assistant-7c-json-base4v2
• Architecture: LlamaForCausalLM
• Approximate size: 5M parameters
• Context length: 512 tokens
• Vocabulary size: 4,108
• Weights format: safetensors
• License: Apache-2.0

Intended Use

Use this model to generate exploratory questions from a topic or short prompt.

Examples:

• cooking
• how to smile
• starting a small business
• database migration risk
• how should I think about learning biology?

The expected output is JSON with a questions array.

Quick Start

import json
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "aungkomyint/tara1.2-quest"

tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(repo_id)
model.eval()

def generate_questions(user_text):
    prompt = f"User: {user_text.strip()}\nAssistant:\n"
    inputs = tokenizer(prompt, return_tensors="pt")
    inputs.pop("token_type_ids", None)

    with torch.no_grad():
        output = model.generate(
            **inputs,
            max_new_tokens=160,
            do_sample=True,
            temperature=0.7,
            top_p=0.9,
            repetition_penalty=1.18,
            pad_token_id=tokenizer.pad_token_id,
            eos_token_id=tokenizer.eos_token_id,
        )

    text = tokenizer.decode(output[0], skip_special_tokens=True)
    reply = text.split("Assistant:", 1)[-1].strip()
    return reply

reply = generate_questions("cooking")
print(reply)

try:
    data = json.loads(reply)
    print("question count:", len(data.get("questions", [])))
except json.JSONDecodeError:
    print("Model did not return valid JSON for this sample.")

Example Output

Prompt:

User: cooking
Assistant:

Sample output:

{
  "questions": [
    "What constraints around time, money, energy, or rules shape cooking?",
    "Which part of cooking is most uncertain right now?",
    "How would a beginner and an expert frame cooking differently?",
    "Why does cooking matter for the larger goal?",
    "What evidence would make your thinking about cooking more reliable?"
  ]
}

Outputs are stochastic when sampling is enabled. Validate JSON in your application.

Training Summary

This checkpoint was trained as a small SFT experiment on top of Tara 1.2 base model work. The goal was to test whether a very small causal language model could learn a narrow JSON question-generation interface.

The training data focused on:

• JSON-formatted question lists
• topic-to-question mapping
• short user prompts
• exploratory and planning-style questions

Evaluation Summary

Local benchmark results showed:

• Strong JSON-format tendency under the expected prompt format.
• Useful simple topic-to-question behavior on common topics.
• Weak semantic grounding on harder prompts.
• Repetition and template overfitting.
• Poor handling of negation and negative constraints.

This release is therefore best treated as an educational/research checkpoint, not a production assistant.

Limitations

• The model is very small, about 5M parameters.
• It can drift off topic.
• It can repeat question templates.
• It may output invalid JSON for some prompts or decoding settings.
• It does not reliably understand negation such as "do not", "not", or "avoid".
• It should not be used for legal, medical, financial, safety, or high-stakes advice.
• It is not designed for multi-turn chat.

Recommended Decoding

Good starting settings:

temperature=0.7
top_p=0.9
repetition_penalty=1.18
max_new_tokens=160

For stricter output, use greedy decoding, but expect more repetition.

Relationship To Tara 1.1

tara1.2-quest is a successor experiment to aungkomyint/tara1.1. Tara 1.1 was a broader tiny assistant experiment. Tara 1.2 Quest narrows the behavior to JSON question generation.

Citation

If you use this model, cite it as:

Aung Ko Myint. Tara 1.2 Quest. 2026. Hugging Face model checkpoint.