To enhance the reasoning capabilities of smaller-sized language models, employ a system of thinking that incorporates mental models, structured Chain-of-Thought processes, and thoughtful reflection before responding to user queries.

Problem:
smaller-sized transformer models exhibit inferior reasoning capabilities compared to their larger counterparts, whose advanced reasoning abilities stem from broader connection networks that facilitate cross-domain inference.

Solution:
Two-Step Approach:

Finetuning: Commence by fine-tuning the Llama 3.1, a smaller-sized transformer model with 8 billion parameters, on an enhanced reasoning dataset to bolster its cognitive capabilities.

Revelation of Internal Processes: Subsequently, leverage a system of thinking model guidance techniques (Think, Plan, Reasoning and Reflection) to unveil the model's internal thought processes and the rationales underlying its processing mechanisms.

Available Mental Models
System of thinking, reasoning and reflection

Chain-of-Thoughts

Thinking Fast and Slow

Critical Thinking

Iceberg Mental Model

Second Order Thinking

***Models***
Base Model: mychen76/Llama-3.1_Intuitive-Thinker https://huggingface.co/mychen76/Llama-3.1_Intuitive-Thinker

Quantized: mychen76/Llama-3.1_Intuitive-Thinker_8B_2309_GGUF https://huggingface.co/mychen76/Llama-3.1_Intuitive-Thinker_8B_2309_GGUF

HF Usage

notebook: https://github.com/minyang-chen/intuitive_thinker/blob/main/intuitive_thinker_demo.ipynb

python

from intuitive_thinker.mental_model import MentalModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch
import json

question="count number of r in word strawberry?"

## format question using mental model template
mental_model = MentalModel(MentalModel.CHAIN_OF_THOUGHTS)
prompt = json.loads(mental_model(question))

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype="float16", bnb_4bit_use_double_quant=True
)
        
# Prepare the input as before
messages = [
    {"role": "system", "content": prompt['system_message']},
    {"role": "user", "content": prompt['user_input'] }
]

hf_model_id = "mychen76/Llama-3.1_Intuitive-Thinker"

# 1: Load the model and tokenizer
model = AutoModelForCausalLM.from_pretrained(hf_model_id, device_map="auto", quantization_config=bnb_config, torch_dtype=torch.bfloat16)
tokenizer = AutoTokenizer.from_pretrained(hf_model_id)

# 2: Apply the chat template
formatted_chat = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
#print("Formatted chat:\n", formatted_chat)

# 3: Tokenize the chat (This can be combined with the previous step using tokenize=True)
inputs = tokenizer(formatted_chat, return_tensors="pt", add_special_tokens=False)
# Move the tokenized inputs to the same device the model is on (GPU/CPU)
inputs = {key: tensor.to(model.device) for key, tensor in inputs.items()}
#print("Tokenized inputs:\n", inputs)

# 4: Generate text from the model
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.1)
#print("Generated tokens:\n", outputs)

# 5: Decode the output back to a string
decoded_output = tokenizer.decode(outputs[0][inputs['input_ids'].size(1):], skip_special_tokens=True)
print("Decoded output:\n", decoded_output)

result

bash

Decoded output:
 <thinking>
a. Briefly analyze the question and outline our approach:
The question asks for the count of the letter 'r' in the word'strawberry'. This is a simple text analysis task that requires examining the word for the target letter.

b. Present a clear plan of steps to solve the problem:
1. Write out the word'strawberry' in its full form.
2. Count the occurrences of the letter 'r' in the word.
3. Provide the final count as the answer.

c. Use a "Chain of Thought" reasoning process:
1. Start by writing out the word: s-t-r-a-w-b-e-r-r-y
2. Now, count the 'r's: r (1), r (2), r (3)
3. We have found all 'r's in the word.

<reflection>
a. Review our reasoning:
We have broken down the task into simple steps and followed through with the counting process.

b. Check for potential errors or oversights:
- We have accounted for all letters in the word.
- We have specifically targeted the letter 'r' for counting.
- We have counted each occurrence of 'r' correctly.

c. Confirm our conclusion:
Yes, our conclusion is correct. The word'strawberry' contains 3 occurrences of the letter 'r'.

</reflection>
</thinking>

<output>
The word'strawberry' contains 3 occurrences of the letter 'r'.
</output>

Ollama.com
https://ollama.com/mychen76/llama3.1-intuitive-thinker

For direct easy to use each mental model has been package on own model package.

Chain-of-Thoughs: [mychen76/llama3.1-intuitive-thinker:chain-of-thoughts.q5]
Thinking Fast and Slow: [mychen76/llama3.1-intuitive-thinker:thinking-fast-n-slow.q5]
Critical Thinking: [mychen76/llama3.1-intuitive-thinker:critical-thinking.q5]
Iceberg Mental Model: [mychen76/llama3.1-intuitive-thinker:iceberg-mental-model.q5]
Second Order Thinking: [mychen76/llama3.1-intuitive-thinker:second-order-thinking.q5]

Samples

Sample: Chain-of-Thoughts

=========== user question ===================

 Patel family who is planning a road trip from Toronto to Vancouver, covering a total distance of 4,400 km. They want to divide the drive into equal parts over 5 days, with each
...  day's drive being 15 km longer than the previous day. If they drive x km on the first day, what is the value of x that will allow them to complete the trip in 5 days?

=========== model Response ===================

<thinking>
To find the value of x, we need to determine how much distance the family needs to cover each day and ensure it adds up to 4,400 km over 5 days.

Let's break down their driving plan:
1. They drive x km on the first day.
2. Each subsequent day's drive is 15 km longer than the previous day.
3. The trip covers a total distance of 4,400 km.

We can express the total distance covered as a sum of distances driven each day:
Day 1: x
Day 2: x + 15
Day 3: (x + 15) + 15 = x + 30
Day 4: (x + 30) + 15 = x + 45
Day 5: (x + 45) + 15 = x + 60

The sum of these distances should equal the total trip distance:
x + (x + 15) + (x + 30) + (x + 45) + (x + 60) = 4,400 km

Let's simplify this equation by combining like terms:
5x + 150 = 4,400 km

<reflection>
Checking our reasoning process:
- We've correctly expressed the distances driven each day.
- The equation we formed is accurate and reflects the problem statement.
- Let's solve for x to confirm our understanding.

Subtracting 150 from both sides of the equation gives us:
5x = 4,250 km

Dividing both sides by 5 yields:
x = 850 km

Our calculation appears sound. This value for x ensures the family will drive equal distances each day, with the last two days being 30 and 45 km longer than the previous 
ones.

</reflection>
</thinking>

<output>
The Patel family should drive x = 850 km on the first day to complete their trip in 5 days, covering an additional 15 km each subsequent day (resulting in drives of 865 km on 
Day 2, 880 km on Day 3, 895 km on Day 4, and 910 km on Day 5).

This problem illustrates how algebra can be used to solve real-world problems involving equal parts and progressive increases. You might find it interesting to explore other 
scenarios where similar mathematical concepts are applied.

Is there anything else you'd like to clarify or any more questions about this type of problem?
</output>

Ollama Usage

https://ollama.com/mychen76/llama3.1-intuitive-thinker

For direct easy to use each mental model has been package on own model package.

Chain-of-Thoughs: [mychen76/llama3.1-intuitive-thinker:chain-of-thoughts.q5]
Thinking Fast and Slow: [mychen76/llama3.1-intuitive-thinker:thinking-fast-n-slow.q5]
Critical Thinking: [mychen76/llama3.1-intuitive-thinker:critical-thinking.q5]
Iceberg Mental Model: [mychen76/llama3.1-intuitive-thinker:iceberg-mental-model.q5]
Second Order Thinking: [mychen76/llama3.1-intuitive-thinker:second-order-thinking.q5]

This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.

Developed by: mychen76@gmail.com
Model type: Llama
License: Follow Llama 3 Licenses
Finetuned from model: Llama3.1 8B

Training Details

Training Data

[More Information Needed]

Training Procedure

Finetuning