SemanticWiki Coder 7B v2 — merged model
What?
This is a standalone Transformers model that generates DeepWiki-style technical documentation from source code. It can produce structured Markdown, source-file references, Mermaid diagrams, tables, and inline citations.
The original fine-tune was a LoRA adapter trained on top of
Qwen/Qwen2.5-Coder-7B-Instruct. This repository contains the adapter merged into the base
weights, so users do not need PEFT or a separate adapter repository.
Why?
The merged format is easier to use with standard Transformers tooling and is the required starting point for conversion to Apple-Silicon MLX format. The merge changes packaging, not the model's intended behavior.
Quick start
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "GhostScientist/semanticwiki-coder-7b-v2-merged"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype="auto",
device_map="auto",
)
messages = [
{"role": "user", "content": "<START_OF_CONTEXT>\n[source code]\n<END_OF_CONTEXT>\n\n<query>\nExplain the architecture.\n</query>"}
]
prompt = tokenizer.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=1024)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Intended use
- Generate first drafts of codebase wikis and architecture notes.
- Keep the supplied source context bounded and review generated citations.
- Use a low temperature for factual documentation and a higher temperature only when exploring alternative explanations.
Limitations and safety
- Citations, diagrams, and explanations can be incorrect or incomplete.
- The model should not be treated as a security auditor or a substitute for human code review.
- Do not include secrets, credentials, or private source code in prompts sent to an untrusted service.
- This model was trained on a small supervised dataset; quality may vary by language, framework, and repository size.
Provenance
- Base model:
Qwen/Qwen2.5-Coder-7B-Instruct - Original adapter:
GhostScientist/semanticwiki-coder-7b-v2 - Fine-tuning method: LoRA supervised fine-tuning
- Merge operation: PEFT
merge_and_unload
License
The base model is distributed under its upstream license. Review the base model card and applicable terms before redistribution or commercial use.