---
base_model:
- Pinkstack/Superthoughts-lite-v1
tags:
- text-generation-inference
- transformers
- unsloth
- llama
- gguf
- cot
- superthoughts
- reasoning
- grpo
license: apache-2.0
language:
- en
datasets:
- openai/gsm8k
- Pinkstack/intructions-sft-sharegpt
---
Demo: https://huggingface.co/spaces/Pinkstack/Chat-with-superthoughts-lite
![superthoughts lite](https://cdn-uploads.huggingface.co/production/uploads/6710ba6af1279fe0dfe33afe/K5kYIHYj2aX2kB6MlcM9O.png)

# Information
Advanced, high-quality and **lite** reasoning for a tiny size that you can run on your phone.

At original quality, it runs at ~400 tokens/second on a single H100 Nvidia GPU from Friendli. 

Trained similarly to Deepseek R1, we used Smollm2 as a base model, then we've SFT fine tuned on reasoning using our own private superthoughts instruct dataset which includes a mix of code, website generation, day-to-day chats, math and counting problems. And then we modified the tokenizer slightly, after the SFT fine tuning we used GRPO to further amplify it's mathematics & problem solving abilities.

# Which quant is right for you?

***F16***: Least hallucinations, high-quality reasoning yet heavy to run.
***Q8_0***: Limited amount of hallucinations high-quality reasoning, recommended
***Q6_k***: Hallucinates more, good reasoning but may fail at counting etc. only use if you cannot run Q8_0.
***Q4_k_m***: Not recommended, Hallucinates, doesn't always think properly. easier to run though.

<div style="background-color: #ffebee; padding: 16px; border-radius: 4px; border-left: 4px solid #ef5350;">
  <h1 style="color: #c62828; margin: 0 0 8px 0;">⚠️ WARNING</h1>
  <p style="color: #c62828; font-weight: bold; margin: 0;">
    We did not put additional safety filters when doing SFT, thus this AI is pretty uncensored and can be rude at times. unless you specify in the system prompt that it is harmless, it won't be. users are soley responsible for the use of this AI. no output from the AI represents the views of Pinkstack or any other third party, it may create biased, inccorect and harmful information unless you set it up properly. for commercial use, it is reccomeneded that you either use another AI like llamaguard for filtering or in the system prompt ensure that it would be harmless.
  </p>
</div>


# Format
```
<|im_start|>user
How many R's in strawberry<|im_end|>
<|im_start|>assistant
<think>
Alright, the user has asked how many R's in the word strawberry, that's easy! I just need to count each instance of the letter 'R' in the word 's-t-r-a-w-b-e-r-r-y' and then find out how many R's there are, lets count!
S - Not an R,
T - Not an R,
R - First instance of the letter R! (1),
A - Not an R,
W - Not an R,
B - Not an R,
E - Not an R,
R - Great! Second instance of the letter R. (2),
R - Third instance of the letter R. (3),
Y - Not an R.

So, i've counted all the letters correctly, meaning that I am sure that there are 3 R's in the word Strawberry. I should probably let the user know.
</think>
<output>3
</output><|im_end|>
```
We reccomend to use a low temperatures as higher values may cause it to not think. 
# system prompt
(important to ensure it would always think, output).
```
respond in the following format:
<think>
...
</think>
<output>
...
</output>
```
# Examples:
all responses below generated with our system prompt and a temperature of 0.7.
Generated inside the android application, ChatterUI via GGUF Q8, using the model's prompt format. and our 
1)
![image/png](https://cdn-uploads.huggingface.co/production/uploads/6710ba6af1279fe0dfe33afe/5veZJmkjuv_7W7pKhvsu0.png)
2)
![image/png](https://cdn-uploads.huggingface.co/production/uploads/6710ba6af1279fe0dfe33afe/pAwPdVkEZ7rnFf-TZ5tMU.png)
3)
![image/png](https://cdn-uploads.huggingface.co/production/uploads/6710ba6af1279fe0dfe33afe/FDaWAAqgv2kvoZvjl8gjl.png)

# Uploaded  model

- **Developed by:** Pinkstack
- **License:** apache-2.0
- **Finetuned from model :** HuggingFaceTB/SmolLM2-1.7B-Instruct