Instructions to use netis-ai/smlr-metrics-1b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use netis-ai/smlr-metrics-1b with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="netis-ai/smlr-metrics-1b", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("netis-ai/smlr-metrics-1b", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("netis-ai/smlr-metrics-1b", trust_remote_code=True)
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=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use netis-ai/smlr-metrics-1b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "netis-ai/smlr-metrics-1b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "netis-ai/smlr-metrics-1b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/netis-ai/smlr-metrics-1b

SGLang

How to use netis-ai/smlr-metrics-1b 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 "netis-ai/smlr-metrics-1b" \
    --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": "netis-ai/smlr-metrics-1b",
		"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 "netis-ai/smlr-metrics-1b" \
        --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": "netis-ai/smlr-metrics-1b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use netis-ai/smlr-metrics-1b with Docker Model Runner:
```
docker model run hf.co/netis-ai/smlr-metrics-1b
```

Netis SMLR Tech Preview — `smlr-metrics-1b`

Netis SMLR Tech Preview. This is a research preview of the Streaming Multi-Lane Reasoner (SMLR). It is provided for evaluation and research; it is not a production monitoring system. Behavior, APIs, and weights may change. Keep a human in the loop for any real operational action.

SMLR is a small language model that watches instead of answering. Instead of "complete input → one answer," it runs as a continuous cognitive process over a live event stream (metrics / logs), deciding at every frame whether to stay silent or escalate — and it is trained to discriminate before it alerts: notice an anomaly, form a hypothesis, query a diagnostic tool, read the decisive field, and only then cry wolf.

This checkpoint is the 1B metrics tier (MiniCPM-derived Llama backbone). Full write-up, architecture, training recipe, evaluation, and a live demo app: https://github.com/Netis/smlr

Architecture

A fused multi-head VLA: one shared backbone + one policy head + six per-lane decode heads.

policy head — Linear(hidden, 10); reads the prompt-end hidden state and predicts next_action (WAIT · NOTE · SUMMARY · QUESTION · VERIFY · QUERY_TOOL · WARN · ALERT · RESOLVE · REVISE) before any lane is decoded.
6 lane heads — observation · reasoning · public_output · notes · state_patch · actions, each decoded in parallel through its own head sharing one prefill.
closed loop — state_patch merges into the working state and re-enters the next frame; actions drives a tool loop.

There is no single lm_head — the lane heads replace it. The model is loaded with trust_remote_code (custom class SmlrMultiLaneForCausalLM in modeling_smlr.py).

Usage

import torch, json
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "netis-ai/smlr-metrics-1b"
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    repo, trust_remote_code=True, dtype=torch.bfloat16
).eval()
model.to("cuda" if torch.cuda.is_available() else "cpu")

SYSTEM = tok.init_kwargs.get("smlr_system") or open("SYSTEM_PROMPT.md").read()  # or paste the system prompt

frame = {
    "recent_window": [], "working_state": {"link_state": {}}, "retrieved_memory": [],
    "new_event": {"seq": 42, "t": 42.0, "type": "metrics_frame",
        "links": {"link0": {"latency_ms": 320, "loss_pct": 7.5, "jitter_ms": 45, "throughput_mbps": 120},
                  "link1": {"latency_ms": 8,   "loss_pct": 0.0, "jitter_ms": 1,  "throughput_mbps": 940}}},
}
prompt = (f"<|im_start|>system\n{SYSTEM}<|im_end|>\n"
          f"<|im_start|>user\n{json.dumps(frame)}<|im_end|>\n<|im_start|>assistant\n")
ids = tok(prompt, return_tensors="pt").input_ids.to(model.device)

out = model.decode_frame(ids)                     # shared prefill + policy + per-lane decode
print("next_action:", out["next_action"])         # e.g. WAIT (first frame; needs sustain to escalate)
for lane, toks in out["lanes"].items():
    if toks:
        print(lane, "→", tok.decode(toks, skip_special_tokens=True)[:200])

decode_frame runs one monitoring frame end to end (a shared prefill, the policy read at prefill latency, then greedy per-lane decode). For real-time monitoring you carry state_patch back into the next frame's working_state — see the demo app.

The exact system prompt used in training/eval ships with the demo app in the GitHub repo.

Intended use & scope

In scope: research on streaming/closed-loop monitoring agents; anomaly detection & alerting on metrics streams in the discriminate-then-alert regime.
Out of scope: a certified production incident system; safety-critical automated remediation without a human; root-cause attribution as sole source of truth.
Domain. Trained on a specific synthetic monitoring domain (network-link metrics; specific host-log incident types). On arbitrary real host telemetry (raw CPU/memory, general syslog) it is out-of-distribution — the demo maps real host signals into the trained schema, and correctly stays quiet on a healthy machine, reacting to genuine sustained load. Treat detections as a preview.

Evaluation (held-out, deliberately zero-shot)

Metric	value
Metrics detect / alert / false-alerts	1.0 / 1.0 / 0
Metrics root-cause	0.75
Logs detect / alert / false-alerts (this 1B tier)	0.75 / 0.75 / 0

Sample sizes are small (n=16 metrics, n=12 logs) — read directional signals only. Full methodology and the research history (including negative results) are in the technical report.

Known limitations

Root cause is weak on the harder cascade metric (~0.44) and is outside the real-time bar.
dns_failure is a stable blind spot; thermal_throttle shows run-to-run flips.
This is the metrics tier; the logs tier is a separate 4B model (not in this release).

License

MIT for the SMLR code and this modeling file. The base model is MiniCPM-derived and remains under its upstream license; downstream users are responsible for upstream license compliance.

Downloads last month: 152

Safetensors

Model size

2B params

Tensor type

BF16