reelva/phi3-mini-reasoning-beast

🧠 Phi-3 Mini Reasoning Beast

The Reasoning Beast is a high-performance fine-tune of Phi-3-mini-4k-instruct optimized for complex cognitive tasks. By integrating the Opus 4.6 Filtered Thinking dataset, this model features native Chain-of-Thought (CoT) capabilities, enabling a structured internal deliberation process before output generation.

🚀 Specifications

Attribute	Details
Architecture	Phi-3-mini-4k-instruct
Parameters	3.8B
Training Hardware	NVIDIA Blackwell (sm_120)
Fine-tuning Dataset	Opus 4.6 Filtered (Thinking/Reasoning focus)
Context Length	4K Tokens
Key Feature	Native <think> tag integration

---

🛠️ Quick Start (Fixes RoPE KeyError)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig, TextStreamer

model_id = "reelva/phi3-mini-reasoning-beast"

# Mandatory RoPE patch for modern Transformers compatibility
config = AutoConfig.from_pretrained(model_id)
if hasattr(config, "rope_parameters"):
    config.rope_parameters["original_max_position_embeddings"] = 4096

model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    config=config, 
    torch_dtype=torch.float16, 
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_id)
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=False)

# Reasoning Prompt Template
prompt = "<|user|>\nSolve this logic puzzle step by step: [Task]<|end|>\n<|assistant|>\n"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

_ = model.generate(
    **inputs, 
    streamer=streamer, 
    max_new_tokens=1500, 
    do_sample=True, 
    temperature=0.7
)

---

⚡ Reasoning Demonstration (Zero-Shot Capability)

The following output demonstrates the model's high-level cognition when encountering an undefined task. Instead of hallucinating, it performs a systematic meta-analysis of the required logical framework.

Input:

<|user|> Solve this logic puzzle step by step: [Task]<|end|>
<|assistant|>

Output:

<|assistant|> <think>
Let me solve this logic puzzle step by step...

First, I need to understand the puzzle by reading through it carefully. The question mentions:

**Logic Puzzle: [Task]
Solve this logic puzzle step by step: [Puzzle Description]**

I see that the puzzle description is missing, so I can't solve the specific puzzle you're asking about.

**My approach:**
1. **Read the puzzle carefully:** Identify the key elements, constraints, and relationships involved.
2. **Create a grid or chart:** This helps visualize the information and make connections between different pieces of data.
3. **Apply logical reasoning:** Use deductive reasoning to eliminate possibilities that don't fit the constraints.
4. **Check each step:** Verify that each deduction is consistent with all the information given.
5. **Review the solution:** Ensure that the solution satisfies all the puzzle's constraints and that no contradictions exist.

**Without the specific puzzle content, I can't provide a step-by-step solution.** If you can provide the details of the puzzle you're working on, I'd be happy to walk through the solution process with you!
</think>

I'm sorry, but it looks like the actual logic puzzle (or "Task") you're asking about isn't provided in the message you've sent.

**Could you please provide the specific logic puzzle you'd like solved?**
<|end|>

---

📚 Citations & References

This model's reasoning capabilities are built upon the foundations laid by the following research:

• Phi-3 Technical Report: Phi-3: A Highly Capable Language Model Locally on Your Phone
• Chain-of-Thought Prompting: Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
• Training Methodology: Inspired by advanced CoT fine-tuning on high-quality synthetic reasoning datasets and the "Abliterated" architectural concepts.

⚠️ Limitations and Usage

• Objective: Optimized for logical reasoning, mathematical deduction, and code analysis.
• Safety: Due to its experimental reasoning stream, standard safety guardrails may behave differently. Use with professional discretion.
• Context: Best performance is achieved using the <|user|> and <|assistant|> prompt structure.

📜 License

This model is released under the MIT License.

---

[ EOF - Logic Protocol Active ]