NexusGraph can be used with local or cloud LLM backends to produce explainable traversal-aware responses. The following Hugging Face model releases and adapters are compatible with our inference pipelines and served as GGUF / adapter artifacts for low-latency inference.
| Model | Type | Hugging Face | Notes |
|---|---|---|---|
| llama3-2-lora-wisdom | LoRA / GGUF | https://huggingface.co/Sandeep4235/llama3-2-lora-wisdom | Fine-tuned Llama-3 2B LoRA weights for concise, citation-first reasoning. Useful for short-context graph trace summarization and low-latency on-device runs via Ollama or llama.cpp. |
| qwen2.5-7b-adapter | Adapter | https://huggingface.co/Sandeep4235/qwen2.5-7b-adapter | Adapter weights for Qwen 2.5 7B which injects compliance-aware templates and token-level suppression for hallucination control. Use with Qwen base models and the adapter loader in your inference stack. |
This snippet shows how to resolve a model call from a FastMCP tool and run it through the local inference client while preserving the NexusGraph verification pipeline.
from backend.inference import get_local_client # hypothetical helper
from backend.mcp_server.tools import analyze_blast_radius_and_commits
client = get_local_client(model_name="Sandeep4235/llama3-2-lora-wisdom")
def run_model_with_trace(prompt, trace_metadata):
# attach the deterministic traversal trace as structured context
context = {
"trace_id": trace_metadata["trace_id"],
"graph_hops": trace_metadata["hops"],
"verifiable_citations": trace_metadata["citations"],
}
response = client.generate(
prompt=prompt,
context=context,
max_tokens=300,
)
# Post-check: ensure no unauthorized data or internal endpoints leaked
sanitized = backend.guardrails.post_check(response)
return sanitized- Model weights & adapters are published on Hugging Face: