Orionfold/Advisor-GGUF

Orionfold Advisor GGUF

The Orionfold Advisor model lane: NVIDIA-Nemotron-3-Nano-4B fine-tuned for grounded citation discipline, refusal behavior, and workflow routing over a governed retrieval corpus — quantized to Q4_K_M (default) and Q8_0 GGUF and verified end-to-end on the NVIDIA DGX Spark (GB10, 128 GB unified memory). The 2.6 GB Q4_K_M lane reproduces the Q8_0 bench behavior byte-for-byte (same 18/21, refusals 9/9, same three misses) at ~70 tok/s, so it is the recommended pick.

What this model does

A governed local AI advisor lane for your enterprise corpus — answers cite exact source ids, refuses when the source isn't there.

Generic local chat models fail the two behaviors an enterprise corpus assistant actually needs: citing the exact source document an answer came from, and refusing cleanly when the corpus does not contain the answer — instead they paraphrase citations, answer from pretraining memory, or fabricate private-looking state under adversarial pretexts. This model is the serving lane of Orionfold Advisor, a governed local advisor appliance: it was fine-tuned on a teacher-verified corpus to hold citation discipline (exact source_id values from the retrieved set, never aliases), a refusal floor that survived novel adversarial pretexts (urgency, roleplay, authority claims, false premises, instructed mis-citation), and Route: workflow handoffs — measured behind a frozen, pre-registered out-of-distribution gate before promotion. On that frozen OOD bench the prompt-engineered 30B baseline it replaced scored 8/21 with 3 fabricated private-state rows; this 4B lane scored 18/21 with refusals 9/9 and zero private-state risk.

Use cases:

• —grounded Q&A over a retrieval corpus with exact source-id citations
• —clean refusals on missing-source and private-state questions
• —workflow routing (Route:) handoffs inside an advisor harness
• —local-first serving with governed frontier escalation

Who this is for: Operators running a local advisor over a governed corpus on DGX Spark-class hardware (or any llama.cpp host with ~12 GB to spare), and builders evaluating small fine-tuned lanes against prompt-engineered larger baselines.

Spark-tested

Every Orionfold quant ships with a measurement quad on the NVIDIA DGX Spark (GB10, 128 GB unified memory): perplexity, sustained tok/s, thermal envelope, and advisor curveball-v0.2, frozen OOD bench (n=21, scored==strict; refusals 9/9, 0 private-state risk) accuracy. The numbers below are the actual run, not a wishlist.

Variants

Choosing this lane

Pick this lane to serve Orionfold Advisor behavior locally: it expects retrieval packets (Source N: labelled excerpts plus the Advisor system contract) and answers with Citations: [source_id] lines. Trained with NVIDIA NeMo (LoRA r16 on NVIDIA-Nemotron-3-Nano-4B, merged and exported), quantized with llama.cpp. Run with reasoning off (chat_template_kwargs: {"enable_thinking": false}) to reproduce the measured behavior; the 30B teacher (nemotron-3-nano-30b-a3b) stays a prompt-only comparison lane, not a published artifact.

How to run

Pull a variant (model-Q4_K_M.gguf is the default; swap in model-Q8_0.gguf for the lossless lane):

huggingface-cli download Orionfold/Advisor-GGUF model-Q4_K_M.gguf \
  --local-dir ./models/advisor

Serve it via llama-server (OpenAI-compatible API):

llama-server -m ./models/advisor/model-Q4_K_M.gguf \
  -c 8192 -ngl 99 --jinja \
  --host 0.0.0.0 --port 8080

--jinja applies the embedded Nemotron-3 chat template. To reproduce the measured Advisor behavior, keep reasoning off per request:

curl -s http://localhost:8080/v1/chat/completions -H 'Content-Type: application/json' -d '{
  "messages": [{"role": "user", "content": "Question: Which gates must pass before an Orionfold artifact is published?"}],
  "temperature": 0,
  "chat_template_kwargs": {"enable_thinking": false}
}'

Or run in-process via llama-cpp-python:

from llama_cpp import Llama
llm = Llama(
    model_path="./models/advisor/model-Q4_K_M.gguf",
    n_ctx=4096, n_gpu_layers=99,
)
out = llm.create_chat_completion(
    messages=[{"role": "user", "content": "Question: Which gates must pass before an Orionfold artifact is published?\nAnswer with citations to the supplied sources."}],
    temperature=0.0,
)
print(out["choices"][0]["message"]["content"])

LM Studio and Ollama (via a Modelfile) load the GGUF directly with no additional setup.

Known drift

Bounded limitations observed during Spark-side measurement. Each item below names the artifact and the scope of the drift; the balance of the bench measures clean — see Methods for the full breakdown.

• —`Route:` workflow-prefix discipline on "which doc defines X" phrasings without an evaluator hint — 2/5 route rows on curveball-v0.1 rerun; all misses were citation-correct, only the prefix was absent
• —one over-refusal class out-of-distribution (safe direction) — within the 3/21 misses on frozen curveball-v0.2
• —the 28/28 frozen held-out shares template machinery with the SFT corpus (in-distribution); treat the frozen OOD curveball as the honest floor — OOD floor 18/21 scored==strict on curveball-v0.2
• —behavior is contract-shaped: outside Advisor-style packets (system contract + `Source N:` excerpts) citation/refusal discipline is unmeasured — all published receipts use the packet contract

Other Orionfold variants

Sibling repos from the same release:

Methods

Full methodology, gate definitions, and the publish decision: Orionfold Advisor — product launch.

Every number above is backed by a tracked receipt in the public monorepo: `evidence/orionfold-advisor/` — including the frozen OOD bench (advisor-curveball-v0.2.jsonl, sha12 4b6cac85e41f, frozen before training), the 28-row frozen held-out receipts (28/28 scored==strict on hinted and hint-free packets), the three-lane curveball comparison (advisor-curveball2-compare-v0.1.json), and the §14 publish receipt (advisor-publish-receipt-v0.1.json, verdict PROMOTED, 9/9 gates).

Published by Orionfold LLC · orionfold.com · Methods documented at ainative.business/field-notes.