Lets try this Mixtral thing because everyone likes MOE right?
Experiemental Mermaid Model: 2x7B Mermaid Mixtral-2x7b
Lets see if its any good, this is 1 epoch of a synthetic dataset exclusively created using my dataset augmentation toolkit using MermaidMistral 7B and MermaidSolar outputs at factual temp range 0.1 to 0.5.
Using my method I have created a dataset of mermaid diagrams from models originally trained by the original 500 hand curated dataset entries at varying temperature ranges.
My toolkit is released for others to expand their dataset with more diverse examples.
This model is an example of training using this method.
Link: https://github.com/Troys-Code/AI_Research/tree/main
Send this as the most simple example of how to use my model for code to mermaid flow diagrams. The rest of the prompt engineering is up to you.
Example to excite the Prompt Engineers out there, Many people have been sending me prompts they use for creating various knowledge graphs, flow diagrams, story board flows, even getting the model to create what if scenario graphs, code flow is its basic skill but it seems like the model is going to keep getting better the more datasets people provide me. Example with something a little more advanced, but please be creative and see what you can get it to do.
The model will auto complete from the word graph TB;
Important Note:
- This is the intution you should understand from how the model likes to perform the best from all my testing so far.
Below you can expect a graph such as this:
graph TB;
A[Start] --> B[Read Input Source];
B --> C{Input Source Type};
C -->|File Path| D[Load File Content];
C -->|String| E[Convert String to List Item];
D --> F[Create Input Data List];
E --> F;
F --> G[Initialize Generator Object];
G --> H[Set Temperature Range];
H --> I[Loop Over Each Prompt];
subgraph Generate Response For Each Prompt
direction TB;
I --> J[Get Current Prompt];
J --> K[Call OpenAI API];
K --> L{Response Unique?};
L --> |No| M[Increase Temperature];
M --> N[Retry With New Temperature];
L --> |Yes| O[Convert to Image];
O --> P[Add Entry to All Entries];
N --> I;
end;
I --> Q[All Prompts Processed];
Q --> R[Write Output to File];
R --> S[End];
style J fill:#ddd,stroke:#777;
style K fill:#ccc,stroke:#777;
style L fill:#eee,stroke:#777;
style M fill:#ff9,stroke:#777;
style N fill:#f99,stroke:#777;
style O fill:#aaf,stroke:#777;
style P fill:#fff,stroke:#777;
Use my toolkit to inference my model and automate some Knowledge Graphs for your own needs.
Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.
Instruction:
Generate the mermaid code block describing the code in excellent detail so I can look at the diagram and understand every single function or the high level diagram in the same full diagram.
Input:
import argparse import json import os import requests import subprocess import tempfile
class MermaidDiagramGenerator: def init(self, theme='dark', background='transparent'): self._theme = theme self._background = background self._entries_dir = os.path.join(os.getcwd(), 'Entries') os.makedirs(self._entries_dir, exist_ok=True)
def convert_to_image(self, mermaid_code, entry_number, output_number):
clean_code = self._remove_mermaid_block_markers(mermaid_code)
output_filename = f"entry_{entry_number}_{output_number}.png"
output_path = os.path.join(self._entries_dir, output_filename)
self._generate_image_from_code(clean_code, output_path)
return output_path
def _remove_mermaid_block_markers(self, code):
code_lines = code.strip().splitlines()
if code_lines[0].startswith("```mermaid") and code_lines[-1] == "```":
return "\n".join(code_lines[1:-1]).strip()
return code
def _generate_image_from_code(self, mermaid_code, output_path):
with tempfile.NamedTemporaryFile(delete=False, mode='w', suffix='.mmd') as temp_file:
temp_file.write(mermaid_code)
input_path = temp_file.name
result = subprocess.run(["mmdc", "-i", input_path, "-o", output_path, "-t", self._theme, "-b", self._background], shell=True, check=False)
os.remove(input_path)
if result.returncode != 0:
raise ValueError("Mermaid diagram generation failed.")
def read_input(input_source): if os.path.isfile(input_source): filename, file_extension = os.path.splitext(input_source) if file_extension == '.json': with open(input_source, 'r') as file: return json.load(file) elif file_extension == '.txt': with open(input_source, 'r') as file: return [{"input": file.read()}] else: return [{"input": input_source}]
def generate_response(prompt, base_temperatures, stream, generator, entry_number, unique_outputs): # prompt_template = f"{prompt}\n\n```mermaid\n"
prompt_template = """
Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.
### Instruction:
Create the mermaid diagram for the following input:
### Input:
{input}
### Response:
```mermaid
""".format(input=prompt)
url = "http://127.0.0.1:5000/v1/completions"
headers = {"Content-Type": "application/json"}
dataset_entries = []
for output_number, temp in enumerate(base_temperatures, start=1):
while True:
data = {
"prompt": prompt_template,
"max_tokens": 4096,
"temperature": temp,
"top_p": 1.0,
"seed": -1,
"top_k": 4,
"repetition_penalty": 1.0,
"guidance_scale": 1.0,
"typical_p": 1.0,
"stream": stream,
}
response = requests.post(url, headers=headers, json=data, verify=False)
response_text = response.json()['choices'][0]['text'].strip()
if response_text.endswith("```"): # Check if response ends with ```
response_text = response_text[:-3].strip() # Remove ``` from the end
if response_text not in unique_outputs:
try:
image_path = generator.convert_to_image(response_text, entry_number, output_number)
print(f"Mermaid diagram generated at: {image_path}")
unique_outputs.add(response_text)
break
except ValueError as e:
print(f"Validation failed, retrying... Error: {e}")
else:
temp += 0.1 # Adjust temperature if output is not unique
dataset_entry = {
"input": prompt,
"output": f"```mermaid\n{response_text}\n```",
"temperature": temp
}
dataset_entries.append(dataset_entry)
return dataset_entries
def generate_unique_responses(input_data, base_temperatures, stream, generator): all_entries = [] unique_outputs = set()
for entry_number, entry in enumerate(input_data, start=1):
prompt = entry.get("input", "")
if prompt:
entries = generate_response(prompt, base_temperatures, stream, generator, entry_number, unique_outputs)
all_entries.extend(entries) # Extend the list with new entries
return all_entries
def main(input_source, stream=False): generator = MermaidDiagramGenerator() input_data = read_input(input_source) base_temperatures = [i / 10 for i in range(5, 11)] # Adjusted for batch of unique outputs per input output_file = "output.json"
all_entries = generate_unique_responses(input_data, base_temperatures, stream, generator)
# Write all entries to the JSON file at once
with open(output_file, "w") as f:
json.dump(all_entries, f, indent=4) # Dump the entire list of entries into the file
if name == "main": parser = argparse.ArgumentParser(description="Generate unique responses and validate Mermaid diagrams.") parser.add_argument('input_source', type=str, help='A multi-line string, path to a .txt file, or a .json file with prompts.') parser.add_argument('--stream', action='store_true', help='Use streaming responses.') args = parser.parse_args()
main(args.input_source, args.stream)
Response:
graph TB;
