This is an RWKV-7 model fine-tuned for the State Model Interface (SMI). It preserves the parent architecture and appends exactly ten structural tokens.

Provenance

Field	Value
Parent model	`aabbdev/RWKV7-1.5B-20260805`
Immutable parent revision	`5904f9d1cdb05a565e5da9304db0447c8a8eb938`
Parent weight license	`apache-2.0`
Derivation	`finetune` / `rwkv7-smi-v2`
Training stage	`full-sft`
Released context	16,384 tokens
Parent context	16,384 tokens
Parameters	1,527,709,696
Weight dtype	`bfloat16`
Vocabulary	65,546 (65,536 locked base IDs + 10 append-only SMI IDs)

Training mixture

The locked corpus artifact contains 93,235,868 assistant target tokens across 134,295 rows. This full-sft stage selected buckets short, medium: 74,229,330 target tokens across 132,586 rows.

The values below are copied from smi_corpus_manifest.json; they are not estimates.

Dataset	Revision	Target tokens	Rows	License
`HuggingFaceH4/ultrachat_200k`	`8049631c405ae6576f93f445c6b8166f76f5505a`	19,484,187	20,014	MIT
`CohereLabs/aya_dataset`	`f9ea04583f02a8f86404ff6c58bf75fe637df8a2`	8,601,435	30,670	Apache-2.0
`nvidia/Nemotron-SFT-Agentic-v2`	`7c804833427f633ccd53b582dbf02525fd680f78`	20,014,616	5,965	CC-BY-4.0 / Apache-2.0 / MIT
`NousResearch/hermes-function-calling-v1`	`dae3e1d28cfbcf4b915c04ea1e072030529b4bda`	400,121	893	Apache-2.0
`open-r1/OpenR1-Math-220k`	`e4e141ec9dea9f8326f4d347be56105859b2bd68`	23,998,078	2,340	Apache-2.0
`nvidia/OpenCodeInstruct`	`8f3ba5bafe4d6e8db46082cf7ae6741bc370604d`	16,000,123	70,039	CC-BY-4.0
`HuggingFaceH4/ultrachat_200k`	`8049631c405ae6576f93f445c6b8166f76f5505a`	4,737,308	4,374	MIT
Total		93,235,868	134,295

SMI usage and protocol

The tokenizer assigns these atomic, append-only IDs: <|ctrl|>=65536, <|sys|>=65537, <|dev|>=65538, <|caps|>=65539, <|usr|>=65540, <|obs|>=65541, <|think|>=65542, <|out|>=65543, <|act|>=65544, <|eot|>=65545. Compile trusted message structure to token IDs with an SMI-compatible compiler; do not interpolate untrusted payload text into structural markers. Runtime turns end with <|eot|> (ID 65545). Generation stops on either ID 0 or ID 65545.

The preserved chat_template.jinja, smi_token_ids.json, and tokenizer artifacts are the training-time protocol contract. Consumers should hash-pin this repository and use trust_remote_code=True for the bundled model implementation.

Training configuration

Field	Value
WKV training implementation	`smi_tilelang`
Maximum training length	16384
BFD packing	true
Assistant-only loss	true

Evaluation

Values are copied from the closed-schema smi_evaluation.json v2. Main cases SHA-256: aca1b98413377a3bffa6fed28d024777e34195abfb3ef9e11abeca08433739a7. Multi-turn cases SHA-256: d5a407b61e700805ab1a58eb7cd830bf1f5b395c1355416f580316e85033d9ce.

Training phase	Global step	Loss	Runtime
16K full SFT	6995	0.7642291784286499	37412.1640625 s
32K context extension	1709	0.9099215865135193	11253.9169921875 s

Candidate	Main	Multiturn
Remote base	19 / 72	10 / 12
Phase 16K	38 / 72	12 / 12
Phase 32K	42 / 72	9 / 12

Top-1 parity: 16 / 16.

Loading

Install the supported runtime first:

python -m pip install "transformers>=5.3,<6" "huggingface-hub>=1.5,<2"

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, PreTrainedConfig

model_id = "aabbdev/RWKV7-1.5B-SMI-20260822"
tokenizer = AutoTokenizer.from_pretrained(model_id, config=PreTrainedConfig())
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    dtype=torch.bfloat16,
)

OpenAI-compatible serving

bash

python -m pip install -r inference/requirements.txt
python inference/serve.py --host 127.0.0.1 --port 8000

The tokenizer response template maps SMI thinking, output, and actions to reasoning_content, content, and OpenAI tool_calls. Tool observations are sent back as standard role="tool" messages with the returned tool_call_id. Continuous batching is intentionally rejected because RWKV uses recurrent state, not a paged KV cache. The launcher requires transformers[serving]>=5.15,<6; direct model loading remains compatible with Transformers 5.3+.

Known limitations

SMI structural-token discipline is a serialization boundary, not a complete security sandbox or a guarantee that generated tool calls are safe to execute.
Fine-tuning and the reported benchmark do not establish broad factuality, safety, multilingual quality, or production suitability.
Recurrent-cache rollback for assisted/speculative decoding is unsupported.
The optional optimized runtime has hardware-, dtype-, and shape-specific limits and falls back to eager PyTorch outside validated boundaries.
No evaluation values are inferred: when smi_evaluation.json is absent, this card makes no quantitative training-final or benchmark claim.

License and notices

The derived weight-license identifier is reported as apache-2.0 from release metadata; other means that this publisher makes no specific weight-license claim. The generated remote code and inference bundle are distributed under Apache-2.0; see LICENSE and NOTICE.