Saul-7B-Instruct-v1-GGUF/README.md

license, library_name, base_model, pipeline_tag, model_creator, language, tags

license	library_name	base_model	pipeline_tag	model_creator	language	tags
mit	gguf	Equall/Saul-7B-Instruct-v1	text-generation	Orionfold LLC	en	gguf spark-tested orionfold base_model:Equall/Saul-7B-Instruct-v1

Saul 7B Instruct v1 GGUF

GGUF quantizations of Equall/Saul-7B-Instruct-v1, verified end-to-end on the NVIDIA DGX Spark (GB10, 128 GB unified memory).

Notebooks

Two runnable notebooks ship with this model — open either on a free cloud GPU:

Notebook	What it does	Open
Builder	Reproduce this model's build and DGX Spark benchmarks end-to-end with fieldkit.
User	Load the published model and call it from your own app in a few lines.

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 LegalBench (n=50, contains) accuracy. The numbers below are the actual run, not a wishlist.

Variant	Size	Perplexity (wikitext-2)	tok/s on Spark	LegalBench (n=50, contains)
Q4_K_M	4.1 GB	5.986	29.4	62.0%
Q5_K_M	4.8 GB	5.938	20.2	72.0%
Q6_K	5.5 GB	5.925	22.4	68.0%
Q8_0	7.2 GB	5.914	7.3	66.0%
F16	13.5 GB	5.917	10.9	68.0%

Thermal envelope: sustained-load minutes before thermal throttle on a single GB10 = 2 min. Beyond this, expect tok/s degradation; the duty-cycle disclosure is per Orionfold's quant-card standard.

Variants

Variant	Recommended use
Q4_K_M	Best balance — fits comfortably in Spark unified memory at 70B; default pick.
Q5_K_M	Higher quality than Q4_K_M with modest size bump.
Q6_K	Near-lossless; recommended if memory headroom allows.
Q8_0	Effectively lossless; reach for this when quality matters more than throughput.
F16	Reference — no quantization. Use only for measurement / baseline.

How to run

Pull a variant:

huggingface-cli download Orionfold/Saul-7B-Instruct-v1-GGUF model-Q5_K_M.gguf \
  --local-dir ./models/saul-7b-instruct-v1

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

llama-server -m ./models/saul-7b-instruct-v1/model-Q5_K_M.gguf \
  -c 4096 -ngl 99 -t 8 \
  --host 0.0.0.0 --port 8080

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

from llama_cpp import Llama
llm = Llama(
    model_path="./models/saul-7b-instruct-v1/model-Q5_K_M.gguf",
    n_ctx=4096, n_gpu_layers=99, chat_format="mistral",
)
out = llm.create_chat_completion(
    messages=[
        {"role": "user",
         "content": "Does the following sentence overrule a previous case? "
                    "Sentence: 'curtman is overruled to the extent it conflicts with evans.'"}
    ],
    temperature=0.0,
)
print(out["choices"][0]["message"]["content"])

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

Methods

Full methodology and Spark-side measurement protocol: Vertical-curator quants on Spark — Saul-7B-Instruct-v1-GGUF + LegalBench mini-eval.

Other Orionfold vertical curators

Same Spark-tested recipe across the curator-on-Spark series:

• finance-chat-GGUF — AdaptLLM finance-chat (Llama-2-7B lineage) for FinanceBench-shaped queries
• SecurityLLM-GGUF — Mistral-based cyber-tuned model with CyberMetric mini-eval gating
• II-Medical-8B-GGUF — Qwen3-8B + DAPO reasoning for MedMCQA-shaped queries

Each card lists its own measurement quad; the headline numbers are recorded as the actual sweep ran, never pre-corrected.

---

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

Want to know when the next Orionfold vertical curator drops? Join the launch list at orionfold.com.