Modelpublic

AMFORGE/sam-v1

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

0likes17downloads

Model Card

SAM — Structured Action Model

SAM is a compact (33.4M params, ~127.4 MB FP32) schema-conditioned model that turns natural language into structured JSON actions across 10 domains: robotics, HTTP/REST, MQTT/IoT, databases, workflows, e-commerce, vehicles, smart home, calendar/email, and filesystem.

Built by AMEFORGE on the in-house SparseMind architecture.

SAM is the successor to [Foros](https://huggingface.co/AMEFORGE/foros-v5.3). Where Foros specialized in robotics ROS-JSON, SAM generalizes the approach to the full agentic / workflow stack while preserving the SparseMind architecture.

TL;DR

The cheap path to reliable JSON for agentic systems:

	Today (LLM API)	With SAM
Output reliability	broken JSON → retry loop	atomic-numeric tokenizer + schema-conditioned
Latency	500–3000 ms	~30–200 ms (CPU)
Cost / 1M calls	$$$$	$0 (offline)
Deployment	API key, cloud, privacy concerns	runs on Jetson, Pi, laptop CPU

Benchmark

Evaluated on the SAM Bench v1 — 200 prompts covering all 10 domains across 5 difficulty tiers (atomic / compound / noisy / long-chain / cross-domain).

(Benchmark not yet run. After training, execute `python sam_benchmark.py` to populate this section.)

Benchmark is fully reproducible — see `sam_benchmark.py` or the `AMFORGE/sam-bench` dataset if published.

Input format (schema-conditioned)

<SCHEMA>{...JSON Schema...}</SCHEMA> <DOMAIN_TAG> <TASK>natural language</TASK> =>

Output: a JSON array of operations conforming to the schema.

Domain tags

<ROS> <HTTP> <MQTT> <DB> <WORKFLOW> <ECOMMERCE> <VEHICLE> <HOME> <CAL> <FILE>

Examples

Input	Output
`<ROS><TASK>move to x=0.5 y=-1.2 z=0.8</TASK> =>`	`[{"op":"move","x":0.5,"y":-1.2,"z":0.8}]`
`<HTTP><TASK>get user 42</TASK> =>`	`[{"op":"http_request","method":"GET","url":"/users/42"}]`
`<MQTT><TASK>publish temp 22 to home/livingroom/temp qos 1</TASK> =>`	`[{"op":"mqtt_publish","topic":"home/livingroom/temp","payload":{"value":22,"unit":"celsius"},"qos":1}]`
`<HOME><TASK>turn on bedroom light at 50% blue</TASK> =>`	`[{"op":"set_light","room":"bedroom","brightness":50,"color":"blue"}]`

Highlights

Property	Value
Architecture	SparseMind (decoder-only)
Parameters	33,400,324 (~33.4M)
Size (FP32)	~127.4 MB (~31.9 MB INT8)
Context length	1024 tokens
Tokenizer	`AMEFORGE/sam_tokenizer` (NexusBPE)
Precision	FP32 (INT8 quantization compatible)
Domains	10 (robotics, HTTP, MQTT, DB, workflow, e-commerce, vehicle, home, calendar, file)
Deployment	CPU, GPU, edge (Jetson, Raspberry Pi)

Quick inference

Use the sam_runtime.py SDK for a clean inference path with optional constrained decoding:

python

from sam_runtime import SAM

sam = SAM.from_hub("AMFORGE/sam-v1")    # downloads weights + tokenizer

result = sam.generate(
    task="get user 42 from api.example.com",
    domain="HTTP",
    schema={"type": "array"},
    mode="guarded",                   # JSON-validated decoding
)

print(result["ops"])
# -> [{"op":"http_request","method":"GET","url":"https://api.example.com/users/42"}]

For OpenAI-compatible tool calling, drop-in replacement:

python

result = sam.tool_call(
    tools=[{...openai-style tool spec...}],
    messages=[{"role": "user", "content": "get me user 42"}],
)

Training

SAM was trained on a large, deterministic multi-domain corpus assembled in-house at AMEFORGE. The corpus covers all 10 supported domains across 5 difficulty tiers (atomic / compound / noisy / long-chain / cross-domain), with paraphrase variation, robustness augmentation, and schema conditioning.

Training was performed on a single GPU using a custom optimizer setup tailored to the SparseMind architecture. Full training methodology and the dataset construction pipeline are kept internal as part of AMEFORGE's IP.

Limitations

—English-only. Multilingual extension is future work.
—Schema-conditioned: best results when a JSON Schema is provided in the prompt.
—Domain set is fixed at 10. New domains require fine-tuning or retraining.
—Numeric atomicity is guaranteed within the production-relevant ranges for each domain. Values outside those ranges fall back to subword encoding.
—Not a chat model — single-turn, structured action generation only.

Citation

bibtex

@misc{sam_2026,
  title  = {SAM: A Compact Schema-Conditioned Structured Action Model
            for Agentic AI},
  author = {AMEFORGE},
  year   = {2026},
  note   = {Built on the SparseMind architecture.
            https://huggingface.co/AMFORGE/sam-v1}
}

Made by AMEFORGE — https://huggingface.co/AMEFORGE