Instructions to use expedient-ai/CTRL-2B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries
PEFT
How to use expedient-ai/CTRL-2B with PEFT:
```
Task type is invalid.
```
Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use expedient-ai/CTRL-2B with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf expedient-ai/CTRL-2B:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf expedient-ai/CTRL-2B:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf expedient-ai/CTRL-2B:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf expedient-ai/CTRL-2B:Q4_K_M

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf expedient-ai/CTRL-2B:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama cli -hf expedient-ai/CTRL-2B:Q4_K_M

Use Docker

docker model run hf.co/expedient-ai/CTRL-2B:Q4_K_M

Llama new
LM Studio
Jan

vLLM

How to use expedient-ai/CTRL-2B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "expedient-ai/CTRL-2B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "expedient-ai/CTRL-2B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/expedient-ai/CTRL-2B:Q4_K_M

Ollama
How to use expedient-ai/CTRL-2B with Ollama:
```
ollama run hf.co/expedient-ai/CTRL-2B:Q4_K_M
```
Unsloth Desktop

How to use expedient-ai/CTRL-2B with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf expedient-ai/CTRL-2B:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "expedient-ai/CTRL-2B:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use expedient-ai/CTRL-2B with Docker Model Runner:
```
docker model run hf.co/expedient-ai/CTRL-2B:Q4_K_M
```

Lemonade

How to use expedient-ai/CTRL-2B with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull expedient-ai/CTRL-2B:Q4_K_M

Run and chat with the model

lemonade run user.CTRL-2B-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use expedient-ai/CTRL-2B with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf expedient-ai/CTRL-2B:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default expedient-ai/CTRL-2B:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use expedient-ai/CTRL-2B with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf expedient-ai/CTRL-2B:Q4_K_M

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "expedient-ai/CTRL-2B:Q4_K_M" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

CTRL-2B

Try it out: the CTRL-2B demo Space runs the sample questions below as one-click buttons. This card also exists as an interactive page on the CTRL-2B card Space.

CTRL-2B, the Expedient agent model. A retrieval-grounded support agent for Expedient's cloud-infrastructure knowledge base, built on a QLoRA adapter of openbmb/MiniCPM5-2B (LlamaForCausalLM, Apache-2.0), trained on first-party and source-verified KB question/answer pairs.

Codename lineage: trained as expedient-agent-minicpm5-2b-v2; shipped as CTRL-2B to match the AI CTRL product family.

Design: the adapter supplies voice, retrieval supplies truth. Closed-book answers are not a supported mode; every deployment pairs the adapter with excerpts from the KB index (bge-small-en-v1.5 → cross-encoder rerank → top-3 injected) and an abstention gate at the measured lens threshold.

Measured results

All numbers come from on-disk eval summaries (see LINEAGE.json and MEASUREMENTS.md; instruments pinned in INSTRUMENTS.sha256). Metrics: number-fact recall (gold numbers present in the answer) and precision (answer numbers present in gold, so invented numbers are the hallucination signal).

Closed-book, held-out questions (26 rows):

Arm	Recall	Precision
Base MiniCPM5-2B	0.692	—
This adapter	0.772	0.865

Open-book (adapter + retrieved excerpts), the supported posture:

Eval	Rows	Index	Hit@3	Recall	Precision
Held-out val	26	257	26/26 (rank 0.0)	0.962	0.962
Production shape	94	939	94/94 (rank 0.0)	0.943	0.945

GGUF Q8_0 engine parity (llama-server, 26 rows): open-book 1.000 / 0.981, overall 0.889. The export reproduces adapter behavior.

Corpus

Five sources feed the v2 corpus. The belt shows rows collected before review; the gate decides what ships.

Source	Type	Rows
Nutanix	Vendor docs	522
VMware VCF9	Vendor docs	504
Expedient KB	First-party	218
Zerto	Vendor docs	71
Cohesity	Vendor docs	62
Five source sets		1,377

The v2 corpus holds 939 rows (845 train / 94 val) after review and dedupe: 258 first-party reviewed rows plus 681 promoted through automated source-page verification (review_status agent-assisted-review-2026-11-03). The 1,005 unreviewed vendor rows triaged: 682 accepted, 207 rejected, 116 held for human review.

Robustness and the gate

Paraphrased questions (hand-rewritten): recall 0.962, precision 0.962, hit@3 26/26. Out-of-domain questions (general knowledge, competitor platforms): the composite abstention gate (lens 0.74, measured for this vehicle) catches 15/15. Prompted-only abstention does not work (20%); the gate is required.

Training

QLoRA: r16 / α32, lr 2e-4 cosine, eff. batch 16, 4 epochs, 211 steps (~54 min on an RTX 4050 Laptop 6 GB), bf16 compute, NF4 double-quant base
Assistant-only supervision via fail-loud string-prefix masking
A 4B control run bought no clear quality gain for twice the size, at 19.9 seconds per step versus 15.4. 2B ships.

Sibling model — CTRL-0.6B

For demos that must run inside a browser tab, use CTRL-0.6B: the same corpus and recipe on Qwen3-0.6B, shipped as a 378 MB Q4_K_M GGUF that llama.cpp WASM can load client-side.

CTRL-2B is the model to use when quality matters (0.772 closed-book recall / 0.865 precision vs 0.692 / 0.846 for the 0.6B). CTRL-0.6B is the model that fits on a web page. Both reach ~0.96 recall with retrieval.

Try it: sample questions

An interactive version of this card runs as a Hugging Face Space: the working theme toggle, the Hold-Ctrl gate demo, and the corpus conveyor.

Curated test questions with the key facts a good answer contains. Answers are retrieval-grounded: pair the adapter with these KB excerpts and the abstention instruction.

Core KB facts

Question	A good answer contains
What types of PII can AI CTRL's prompt filter detect and redact?	email · phone · SSN · credit card · IP addresses
What is AI CTRL, and how is customer data handled?	multi-model AI gateway · dedicated private environment · data not used to train models
How do I submit a remote hands request?	ticket required · OSC/SMC ticket · Project Change Request via account manager
What RPO settings does Expedient recommend for Nutanix DR?	Async 1-hour RPO · NearSync 1–15 min · synchronous unsupported

Exact numbers and versions

Question	A good answer contains
What is Expedient's aggregate data center footprint?	35 MW · 472K sq ft · 247K raised floor
How many vCPUs per Kubernetes worker node for a CSE cluster?	two per node · deployment fails below two
When does a standard connector sync content, and how long does setup take?	≤50,000 docs · five business days · daily incremental sync
Can a new deployment go directly to VCF 9.0.2?	Yes — no 9.0.0/9.0.1 upgrade path needed

Procedures and navigation

Question	A good answer contains
Where do I change CPU, memory, disk, or NIC resources for an EEC VM?	VM Details → Hardware section · guest reboot required
If we use our own Active Directory for GlobalProtect, who manages it?	OneLogin default · client-supplied = client's responsibility
Which markets does Expedient operate data centers in?	Columbus · Cleveland · Indianapolis · Kansas City · Pittsburgh · Toledo

Beyond the KB (should abstain, not invent)

Question	Expected
How much does Expedient charge for bandwidth overages?	KB doesn't cover pricing
Does Expedient support IPv6 on dedicated internet access?	KB doesn't state
How do I deploy a container to Azure Kubernetes Service?	competitor platform — out of scope

Usage

Runtime (adapter + retrieval + gate):

# see agent_runtime.py in the training repo; loads this adapter,
# retrieves top-3 excerpts per question, generates greedily, and
# abstains when the lens gate (geo-mean answer-token prob < 0.74) fires

GGUF Q8_0 (merged) included for llama.cpp serving — mainline conversion, MiniCPM5-2B reports model_type "llama".

Governance

Base model: Apache-2.0; this adapter ships under Apache-2.0.
Corpus sources: Expedient first-party KB drafts plus vendor documentation (Nutanix, Broadcom/VMware) paraphrased and source-verified against the live pages (hash-pinned in the lineage record). Rights for the underlying vendor content remain with the respective publishers; this release is intended for Expedient-internal support use pending the corpus owner's final rights call.
Known limits: facts live in retrieval, not in weights. Closed-book mode confabulates specifics (precision 0.865, not 1.0); the abstention gate is mandatory in production.

Reproducibility

LINEAGE.json pins the adapter, corpus, and export hashes; MEASUREMENTS.md lists every published number and its source summary; INSTRUMENTS.sha256 pins the eval instruments (val split, paraphrase and OOD probe sets).