---
license: cc-by-nc-4.0
language:
- en
library_name: transformers
pipeline_tag: text-classification
base_model: vrishv/Viveka-34M
tags:
- text-classification
- reasoning-detection
- small-model
- tiny-model
- llama
- onnx
- english
- 34m
datasets:
- allenai/ai2_arc
- google/Synthetic-Persona-Chat
- vrishv/jigyasa-simple
- vrishv/jigyasa-medium
- vrishv/jigyasa-big
---

# thinker-34M

**Does this question need careful thinking?** thinker-34M answers `quick` or `think` for a question, so an assistant
can switch a model's slower "thinking" mode on only when it helps. It is [Viveka-34M](https://huggingface.co/vrishv/Viveka-34M)
fine-tuned as a classifier and runs in milliseconds on a laptop CPU, next to [router-34M](https://huggingface.co/vrishv/router-34M).

**Status: experimental.** It is right on **68%** of held-out questions (50% would be guessing),
so use it only when it is very sure. In Vrishvardhan's offline assistant, thinking is switched on only when its
`think` score is 0.85 or more.

| Question | Label |
|---|---|
| hello, how are you? | quick |
| Why is the sky blue? | quick |
| What is the derivative of x squared? | think |

## Use

```python
from transformers import pipeline
clf = pipeline('text-classification', model='vrishv/thinker-34M')
print(clf('What is the derivative of x squared?'))
```

An ONNX copy (`model.onnx`) is included for CPUs without PyTorch, e.g. with `onnxruntime` and `tokenizers`:

```python
import numpy as np, onnxruntime as ort
from tokenizers import Tokenizer
from huggingface_hub import hf_hub_download
tok = Tokenizer.from_file(hf_hub_download('vrishv/thinker-34M', 'tokenizer.json'))
sess = ort.InferenceSession(hf_hub_download('vrishv/thinker-34M', 'model.onnx'))
ids = tok.encode('What is the derivative of x squared?').ids[:128]
logits = sess.run(None, {'input_ids': np.array([ids]), 'attention_mask': np.ones((1, len(ids)), dtype=np.int64)})[0]
print(['quick', 'think'][int(logits.argmax())])
```

## Training

Fine-tuned on Kaggle (two T4 GPUs, two settings at once, the better one kept: learning rate 1e-4, 4 epochs).

| Label | Examples from |
|---|---|
| `quick` | AI2 ARC-Easy (CC BY-SA 4.0), [Jigyasa Simple](https://huggingface.co/datasets/vrishv/jigyasa-simple) (Apache 2.0), small talk from Google Synthetic-Persona-Chat (CC BY 4.0) |
| `think` | AI2 ARC-Challenge (CC BY-SA 4.0), [Jigyasa Medium](https://huggingface.co/datasets/vrishv/jigyasa-medium) and [Big](https://huggingface.co/datasets/vrishv/jigyasa-big) (Apache 2.0) |

## Limitations

- ARC's Easy and Challenge sets are split by whether simple computer methods could answer a question, not by how
  the question reads, so many questions look alike. That limits how well wording alone can tell them apart.
  A better-labelled version is planned.
- English only. Questions are read up to 128 tokens.

## About the maker

thinker-34M was built by **Vrishvardhan Sharma** ([vrishv](https://huggingface.co/vrishv)), a 9-year-old self-taught
builder from India. His other work:

- **GitaVardhan**: a free, offline Bhagavad Gita app in 13 Indian languages, and a pocket-sized voice-controlled
  Gita device he designed, wired and coded himself.
- **Sanskrit Vardhan**: an app for learning Sanskrit through the Bhagavad Gita, with word-by-word meanings and
  step-by-step Pāṇinian derivations. Launched on 20 October 2026 by Prof. Shrinivasa Varakhedi, Vice-Chancellor of
  Central Sanskrit University.
- **Kannada, Tamil and Telugu Vardhan**: apps that teach everyday spoken Kannada, Tamil and Telugu.
- **Hindi Vardhan** ([hindivardhan.in](https://hindivardhan.in)): a dictionary of pure, Sanskrit-rooted Hindi. Over
  6,000 borrowed words in everyday Hindi, each with native alternatives checked against five classical Sanskrit
  dictionaries.
- **AksharaVardhan**: a handwriting reader that runs on an ordinary computer.
- **[graphics-studio](https://huggingface.co/spaces/vrishv/graphics-studio)**: a free Space that makes and edits images, videos, music and narration.
- **[Viveka-34M](https://huggingface.co/vrishv/Viveka-34M)**: his small language model, trained from scratch, that this model is built on.
- **[Shodhana-34M](https://huggingface.co/vrishv/Shodhana-34M)**: Viveka trained further to correct lines misread
  by a handwriting reader.
- **[Jigyasa Simple](https://huggingface.co/datasets/vrishv/jigyasa-simple)**: 473 everyday science questions with
  short, plain-text answers.
- **[Jigyasa Medium](https://huggingface.co/datasets/vrishv/jigyasa-medium)** and **[Jigyasa Big](https://huggingface.co/datasets/vrishv/jigyasa-big)**:
  969 and 970 deeper science questions, built from his own questions, with answers of up to 7 and 10 sentences.
- **[router-34M](https://huggingface.co/vrishv/router-34M)**: picks which of four offline models should answer a question; used next to this thinker.

He has memorised the Bhagavad Gita and Pāṇini's Aṣṭādhyāyī, won national Sanskrit recitation competitions, and
speaks six languages. More at [vrishvardhan.in](https://vrishvardhan.in).

CC BY-NC 4.0 © 2026 Vrishvardhan Sharma.