GlorryLlama-3B-Reasoning

• Developed by: whitelotus0
• License: Apache-2.0
• Finetuned from model: unsloth/Llama-3.2-3B-Instruct

Model Description

GlorryLlama-3B is a fine-tuned version of Llama 3.2 3B, specialized in reasoning and stream-of-consciousness thinking.

It was trained on the ServiceNow-AI/R1-Distill-SFT dataset, which encourages the model to "think" before it answers. The model mimics a reflective assistant that explores, doubts, and refines its own logic before providing a final solution.

This model was trained 2x faster with Unsloth and Huggingface's TRL library.

Intended Use & Prompt Format

To get the reasoning behavior (Chain of Thought), you must wrap your input in the specific <problem> tags and use the system prompt below.

System / Instruction Prompt

You are a reflective assistant engaging in thorough, iterative reasoning, mimicking human stream-of-consciousness thinking. Your approach emphasizes exploration, self-doubt, and continuous refinement before coming up with an answer.
<problem>
{YOUR QUESTION HERE}
</problem>

How to Use

Using Unsloth (Recommended for speed)

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "whitelotus0/glorryllama",
    max_seq_length = 2048,
    dtype = None,
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)

# Define the prompt structure
sys_prompt = """You are a reflective assistant engaging in thorough, iterative reasoning, mimicking human stream-of-consciousness thinking. Your approach emphasizes exploration, self-doubt, and continuous refinement before coming up with an answer.
<problem>
{}
</problem>
"""

# Format the query
query = "How many 'r's are present in 'strawberry'?"
formatted_message = sys_prompt.format(query)

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

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

outputs = model.generate(
    input_ids = inputs,
    max_new_tokens = 1024,
    use_cache = True,
    temperature = 1.5,
    min_p = 0.1
)
print(tokenizer.batch_decode(outputs)[0])

Using Transformers

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("whitelotus0/glorryllama")
model = AutoModelForCausalLM.from_pretrained("whitelotus0/glorryllama", device_map="auto")

prompt_template = """You are a reflective assistant engaging in thorough, iterative reasoning, mimicking human stream-of-consciousness thinking. Your approach emphasizes exploration, self-doubt, and continuous refinement before coming up with an answer.
<problem>
{}
</problem>
"""

text = prompt_template.format("Explain logic clearly.")
inputs = tokenizer(text, return_tensors="pt").to("cuda")

outputs = model.generate(**inputs, max_new_tokens=1024)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Details

Hyperparameters

• Training Steps: 100
• Learning Rate: 1e-4
• Batch Size: 2 (per device) with 4 gradient accumulation steps
• Optimizer: AdamW 8-bit
• Precision: bfloat16
• Quantization: 4-bit (QLoRA)

Dataset

Trained on the ServiceNow-AI/R1-Distill-SFT dataset, which contains reasoning traces generated by DeepSeek-R1.