README.md

metadata

base_model: deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
library_name: transformers
tags:
  - reasoning
  - chain-of-thought
  - deepseek
  - qwen
  - gguf
  - bnb
  - 4bit
  - unsloth
  - chatml
  - agent
  - code
  - thinking
  - distilled
license: apache-2.0

🌌 DeepLink-R1

DeepLink-R1 is a highly specialized, reasoning-focused large language model designed to act as a "Logical Architect." Built on top of the deepseek-ai/DeepSeek-R1-Distill-Qwen-7B architecture, this model doesn't just provide answers—it constructs transparent, mathematically rigorous blueprints of thought.

It is trained to "think" before it speaks using <think> tags, exposing its internal logical deduction process before delivering a final, refined response.

Created by 3amthoughts.

⚡ Model Highlights

• Architecture: 7B Parameters (Base: deepseek-ai/DeepSeek-R1-Distill-Qwen-7B)
• Format: Available in BF16/FP16 (Transformers) and GGUF (Q4_K_M for local execution via llama.cpp/Ollama)
• Capabilities: Deep logical reasoning, mathematical rigor, complex problem-solving, and transparent thought generation.
• Training: Fine-tuned using QLoRA via Unsloth on a high-fidelity mixture of elite reasoning datasets.

🎯 Key Capabilities

• Transparent Reasoning: Trained to output its internal thought processes inside <think> tags, drastically reducing hallucinations and logical leaps.
• High-Tier Logic: Excels at math, coding, and multi-step deduction by breaking down complex problems into manageable architectural blueprints.
• Lightweight & Fast: Distilled into a 7B parameter footprint, making elite-level reasoning runnable on consumer hardware (MacBooks, RTX 3090/4090s).

---

🛠️ How it Works (Prompting & Output)

To get the model to trigger its reasoning engine, you must use the standard ChatML template and prompt it to think step-by-step.

System Prompt:

"You are a logical architect. Think step-by-step."

Example Interaction User:

"If I have a 5-liter jug and a 3-liter jug, how can I measure exactly 4 liters of water?"

DeepLink-R1 Output:

<think>
To measure exactly 4 liters using a 5-liter jug (J5) and a 3-liter jug (J3):
1. Fill J5 to the top (J5 = 5L, J3 = 0L).
2. Pour water from J5 into J3 until J3 is full. This takes 3 liters. (J5 = 2L, J3 = 3L).
3. Empty J3 completely. (J5 = 2L, J3 = 0L).
4. Pour the remaining 2 liters from J5 into J3. (J5 = 0L, J3 = 2L).
5. Fill J5 to the top again. (J5 = 5L, J3 = 2L).
6. Pour water from J5 into J3 until J3 is full. Since J3 already has 2 liters, it only takes 1 liter to fill it. (J5 = 4L, J3 = 3L).
Now, J5 contains exactly 4 liters.
</think>
To measure exactly 4 liters, follow these steps:
1. Fill the 5-liter jug completely.
2. Pour water from the 5-liter jug into the 3-liter jug until it is full, leaving exactly 2 liters in the 5-liter jug.
3. Empty the 3-liter jug.
4. Pour the 2 liters from the 5-liter jug into the empty 3-liter jug.
5. Refill the 5-liter jug completely.
6. Carefully pour water from the 5-liter jug into the 3-liter jug until the 3-liter jug is full. Since it already contains 2 liters, it will only take 1 liter to fill.
You now have exactly 4 liters of water remaining in the 5-liter jug.

💻 Prompt Format (ChatML) DeepLink-R1 strictly uses the ChatML prompt format.

Text
<|im_start|>system
You are a logical architect. Think step-by-step.<|im_end|>
<|im_start|>user
How many 'r's are in the word strawberry?<|im_end|>
<|im_start|>assistant
<think>
...
</think>
...<|im_end|>

🚀 Usage Using transformers (Python)

Python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "3amthoughts/DeepLink-R1"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are a logical architect. Think step-by-step."},
    {"role": "user", "content": "How many 'r's are in the word strawberry?"}
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to("cuda")
outputs = model.generate(inputs, max_new_tokens=1024, temperature=0.6)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))