Text Generation
Transformers
Safetensors
granite
cybersecurity
defensive-security
infrastructure-as-code
dora
rslora
canadian-ai
simpledirect
vinci
conversational
Instructions to use simpledirect/Vinci-Cyber-30B-1.0 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use simpledirect/Vinci-Cyber-30B-1.0 with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="simpledirect/Vinci-Cyber-30B-1.0") messages = [ {"role": "user", "content": "Who are you?"}, ] pipe(messages)# pip install -U transformers accelerate # Load model directly from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("simpledirect/Vinci-Cyber-30B-1.0") model = AutoModelForCausalLM.from_pretrained("simpledirect/Vinci-Cyber-30B-1.0", device_map="auto") messages = [ {"role": "user", "content": "Who are you?"}, ] inputs = tokenizer.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=256) print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:])) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use simpledirect/Vinci-Cyber-30B-1.0 with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "simpledirect/Vinci-Cyber-30B-1.0" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "simpledirect/Vinci-Cyber-30B-1.0", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/simpledirect/Vinci-Cyber-30B-1.0
- SGLang
How to use simpledirect/Vinci-Cyber-30B-1.0 with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "simpledirect/Vinci-Cyber-30B-1.0" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "simpledirect/Vinci-Cyber-30B-1.0", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "simpledirect/Vinci-Cyber-30B-1.0" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "simpledirect/Vinci-Cyber-30B-1.0", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use simpledirect/Vinci-Cyber-30B-1.0 with Docker Model Runner:
docker model run hf.co/simpledirect/Vinci-Cyber-30B-1.0
Commit History
25 days ago
-
style(brand): add restrained Vinci and Vero card identity
Verified signature28963c9
30 days ago
-
eval: add vLLM 0.27.1 output-identity results
Verified signature3cbea4a -
eval: add Vinci Cyber family comparison (8B / 30B / 123B vs parents, same held-out IaC evaluation and HumanEval harness)
Verified signaturef8de7bc
-
docs(model-card): add approved repair and restraint figure
Verified signature14e2ca3 -
docs(model-card): add approved workflow and scoped evidence figures
Verified signature57a3165 -
docs(model-card): apply approved Cyber 30B banner and remove duplicate logo
Verified signature46b45d5 -
docs(model-card): add approved Vinci Cyber 30B banner
Verified signature7a8e7fb -
Pre-release card review: fix lineage metadata, remove internal status text, align sibling claims
George Pu +1: Claude Opus 5 (1M context)- Correct the parent-baseline framing so the raw 0/8 carries its output-contract qualification nearby, per the 2026-09-18 publication ruling. - Do not attribute the 30B parent's format-normalised 2/8 to the 8B parent; the 8B parent has no such re-score and its floor is unmeasured. - Remove pre-publication internal text: private-repo status marker, draft correction history, internal authorization ledger, unreachable-record citations. - Fix front-matter lineage on the 30B GGUF repo (was pointing at IBM Granite rather than Vinci-Cyber-30B-1.0), and align tags across the family. - Add the missing LICENSE and NOTICE where Apache-2.0 4(a)/4(b) require them. - NOTICE files no longer reference a -LoRA repository that does not exist. - Propagate the retracted GSM8K reasoning fix to the 30B GGUF card and correct 0.904 to 0.905; explain the per-tier disclosure split between GGUF siblings. - Cross-link the four repositories; refresh evidence dates. Co-Authored-By: Claude Opus 5 (1M cont…6460019 -
Replace invalid repeat-run reasoning with pre-registered confirmatory result
Verified signatureIdentical greedy repeat scores measure the harness, not the models. Replaced with the pre-registered confirmatory outcome: ARC-Challenge established (+1.06 pp, p=0.014, 1418 held-out items, Holm over four hypotheses), HellaSwag null and bounded below 0.14 pp, no established regression on any task. No table numbers changed.a944a0f -
Card: GSM8K null result now rests on a paired exact-McNemar test, not range overlap
Verified signatureThe range-overlap basis is superseded: under greedy decoding on a fixed item set repeated runs reproduce almost exactly, so overlapping ranges show reproducibility rather than a real difference. Paired exact McNemar over the 1,319 shared questions: 18 disagreements (7 base-right/model-wrong, 11 the other way), paired difference +0.00303, p = 0.48. Conclusion unchanged.6344308
-
Fix GSM8K exclusion justification (range overlap, six runs/arm) and pin base revision in benchmark methods
Verified signatureCHANGE 1: the earlier revision excluded GSM8K by comparing one arm's run-to-run spread to the gap between models, which is the wrong test. Conclusion unchanged; now justified by overlap of the two arms' ranges over six 5-shot runs per arm. CHANGE 2: benchmark methods now name the base arm at revision 4fae6278f7132abf5e971f9de49ebbad09c54cce, matching the provenance pin.dec9911 -
Add general-capability benchmark section (lm-eval-harness 0.4.11, 2026-09-20)
Verified signatureInserts one new section after "## The result": five 0-shot benchmarks vs the base model, plus GSM8K reported as a range. Insertion only; no existing line changed.72e5cee
-
Review finding F5: GGUF validation is complete, not pending; drop the retired degradation-ordering criterion
Verified signaturee9e382e -
Bind provenance to the merged release record (#638 e569e2ae, #644 387318e4) and the publication predicates
Verified signaturee25c678 -
Post-freeze reconciliation: explicit publication/scorer predicates, raw checker clears, and the parent truncation bound
Verified signatured6aa61b -
Owner ruling 2026-09-18 s6: invalid clears travel with the 5/8 in both the headline and the 8B comparison
Verified signature680b5e3
-
Stage Gate-C-qualified merged weights (private staging, not a release)
Verified signature449e624 -
Gate D finding 2: carry the format-normalised bound inline in the comparison table, not only in the headline section
Verified signature19e7cd3 -
Staging: generation_config.json (write probe)
Verified signaturebd91161 -
Lead with the result as a table, including the format-normalised parent column; remove the duplicate summary it replaced
Verified signatured3e3d9c -
Style title as Vinci-Cyber-30B-1.0, add tagline, drop the release-status callout, and clear the stale withholding language it left behind
Verified signature8ff0e1c -
Required caveat: the parent 0/8 is partly an output-contract failure; format-normalised it recovers 2/8, so the honest gain is 2 to 5, not 0 to 5
Verified signature88b747a -
Remove staging scaffolding from the release payload
Verified signatureThese were written by a staging seat before the model card existed. All seven are now either superseded by README.md / MODEL-PROVENANCE.json, stale, or carry internal cross-experiment lineage that should not ship in a public payload -- other seeds' and D206 training job ids, and the D206 corpus digest. GGUF-CONVERSION-STATUS.md said 'NOT PERFORMED', which is false: the merge passed its gate and the GGUFs are built. LICENSE-RIGHTS-CHECKLIST.md was a draft with every box open; rights are settled. Archived outside the repo; the research lineage lives in the research record.072215c -
Remove staging scaffolding from the release payload
Verified signatureThese were written by a staging seat before the model card existed. All seven are now either superseded by README.md / MODEL-PROVENANCE.json, stale, or carry internal cross-experiment lineage that should not ship in a public payload -- other seeds' and D206 training job ids, and the D206 corpus digest. GGUF-CONVERSION-STATUS.md said 'NOT PERFORMED', which is false: the merge passed its gate and the GGUFs are built. LICENSE-RIGHTS-CHECKLIST.md was a draft with every box open; rights are settled. Archived outside the repo; the research lineage lives in the research record.8b289ef -
Remove staging scaffolding from the release payload
Verified signatureThese were written by a staging seat before the model card existed. All seven are now either superseded by README.md / MODEL-PROVENANCE.json, stale, or carry internal cross-experiment lineage that should not ship in a public payload -- other seeds' and D206 training job ids, and the D206 corpus digest. GGUF-CONVERSION-STATUS.md said 'NOT PERFORMED', which is false: the merge passed its gate and the GGUFs are built. LICENSE-RIGHTS-CHECKLIST.md was a draft with every box open; rights are settled. Archived outside the repo; the research lineage lives in the research record.5303b7d -
Remove staging scaffolding from the release payload
Verified signatureThese were written by a staging seat before the model card existed. All seven are now either superseded by README.md / MODEL-PROVENANCE.json, stale, or carry internal cross-experiment lineage that should not ship in a public payload -- other seeds' and D206 training job ids, and the D206 corpus digest. GGUF-CONVERSION-STATUS.md said 'NOT PERFORMED', which is false: the merge passed its gate and the GGUFs are built. LICENSE-RIGHTS-CHECKLIST.md was a draft with every box open; rights are settled. Archived outside the repo; the research lineage lives in the research record.13c5d9f -
Remove staging scaffolding from the release payload
Verified signatureThese were written by a staging seat before the model card existed. All seven are now either superseded by README.md / MODEL-PROVENANCE.json, stale, or carry internal cross-experiment lineage that should not ship in a public payload -- other seeds' and D206 training job ids, and the D206 corpus digest. GGUF-CONVERSION-STATUS.md said 'NOT PERFORMED', which is false: the merge passed its gate and the GGUFs are built. LICENSE-RIGHTS-CHECKLIST.md was a draft with every box open; rights are settled. Archived outside the repo; the research lineage lives in the research record.d27589d -
Remove staging scaffolding from the release payload
Verified signatureThese were written by a staging seat before the model card existed. All seven are now either superseded by README.md / MODEL-PROVENANCE.json, stale, or carry internal cross-experiment lineage that should not ship in a public payload -- other seeds' and D206 training job ids, and the D206 corpus digest. GGUF-CONVERSION-STATUS.md said 'NOT PERFORMED', which is false: the merge passed its gate and the GGUFs are built. LICENSE-RIGHTS-CHECKLIST.md was a draft with every box open; rights are settled. Archived outside the repo; the research lineage lives in the research record.3b268a5 -
Remove staging scaffolding from the release payload
Verified signatureThese were written by a staging seat before the model card existed. All seven are now either superseded by README.md / MODEL-PROVENANCE.json, stale, or carry internal cross-experiment lineage that should not ship in a public payload -- other seeds' and D206 training job ids, and the D206 corpus digest. GGUF-CONVERSION-STATUS.md said 'NOT PERFORMED', which is false: the merge passed its gate and the GGUFs are built. LICENSE-RIGHTS-CHECKLIST.md was a draft with every box open; rights are settled. Archived outside the repo; the research lineage lives in the research record.6a9c1ce -
Publish the internal IaC results: full per-case table, the head-to-head against 8B, and the failures — with the scorer status stated rather than the numbers withheld
Verified signature573dc16 -
Add 8B/30B family comparison and which-model-to-use guidance; state plainly that the head-to-head capability comparison is not yet drawable
Verified signatureeaa4aa1 -
Add Vinci logo (light/dark pair, matching the convention used on other Vinci model cards) and attribute development to SimpleDirect, a Canadian AI Lab
Verified signature2fc6da5 -
Add LICENSE (release payload)
Verified signature5f8aedf -
Add MODEL-PROVENANCE.json (release payload)
Verified signature2b26711 -
Model card: replace staging skeleton with the written release card; refresh stale rights and format claims
Verified signatureThe skeleton said 'Rights cleared - NOT MET', carried license: other / pending-owner-decision, and tagged the repo lora/peft/not-for-release. All four are stale: the parent licence is established Apache-2.0 from the pinned revision's own README, the corpus-rights question is closed by owner ruling, and the distributed artifact is full merged weights, not an adapter. Weights are still NOT in this repo and the card says so explicitly, so its loading instructions are not mistaken for a working path. MODEL-PROVENANCE.json carries the SHA-256 of every file that will land, written before the upload.439f657 -
Upload folder using huggingface_hub
Verified signatured9119e2 -
initial commit
Verified signaturef246fad