StandardThinking/StandardOne-3B-LoRA

Standard One 3B (LoRA adapter)

Updated weights (v2.1, 2026-09-27). If you downloaded this adapter before, download it again. Earlier releases remain available under the v1, v1.1 and v2 tags.

Version: v2.1

Standard One scores a bounded set of answers for a supplied scenario and returns probabilities through POST /v1/systemone. It does not generate free-form response text. This repository contains the 3B LoRA adapter and a merge recipe. Serving requires a merged checkpoint and the server code in StandardOne-8B.

Image input is supported; the results below measure text decisions.

At a glance

• choice selects among labeled options, noul is yes/no and score uses an ordinal scale.
• The training mixture covers English, Japanese, Chinese, Spanish, French, German, Portuguese, Russian and a smaller Korean share. Performance varies by language.
• Probabilities are fitted by answer type; the measured endpoint configuration and calibration results are below.
• The base, adapter, merged weights and shared server code are Apache-2.0.

Repositories

• Merged BF16 checkpoint
• LoRA adapter and merge recipe
• Shared server code

Merge the adapter

import torch
from transformers import Mistral3ForConditionalGeneration
from peft import PeftModel

base = Mistral3ForConditionalGeneration.from_pretrained(
    "mistralai/Ministral-3-3B-Instruct-2512-BF16",
    revision="b6d637bef2393152b3da2b2fde72eecdee30557e",
    torch_dtype=torch.bfloat16,
)
model = PeftModel.from_pretrained(base, ".").merge_and_unload()
model.save_pretrained("./StandardOne-3B-merged", safe_serialization=True)
# Copy tokenizer, chat template, processor and generation files from the base snapshot.

Quick start

Use the merged StandardOne-3B checkpoint or merge the adapter above. Follow the server setup on a CUDA-capable Linux host to install SGLang 0.5.20 and the adapter in separate virtual environments. Start the engine and the adapter in separate terminals. Use the 3B commands below; the linked guide also covers another model size.

Engine:

CUDA_VISIBLE_DEVICES=0 SGLANG_VLM_CACHE_SIZE_MB=0 .venv-sglang/bin/python -m sglang.launch_server \
  --model-path ./StandardOne-3B-merged --served-model-name standard-one-3b \
  --host 127.0.0.1 --port 30000 --tp-size 1 --model-impl sglang --dtype bfloat16 \
  --context-length 8192 --max-running-requests 32 --mem-fraction-static 0.8 \
  --chunked-prefill-size -1 --disable-radix-cache --mm-preprocess-cache-size-mb 0 \
  --model-config-parser hf --load-format safetensors

Adapter:

.venv-native/bin/jev-adapter --engine-url http://127.0.0.1:30000 --model standard-one-3b --alias jev-latest \
  --host 0.0.0.0 --port 30120 --max-concurrency 1 \
  --tokenizer-model mistralai/Ministral-3-3B-Instruct-2512-BF16 \
  --tokenizer-revision b6d637bef2393152b3da2b2fde72eecdee30557e \
  --prompt-wording native --native-system-prompt none --default-temperature 0.90 --temperature-by-type choice=0.90,noul=1.15,score=1.20

The endpoint accepts choice (labeled options), noul (yes/no) and score (ordinal) questions. It returns probabilities for the caller's labels in one forward pass without decoding answer text. For example:

curl -s http://127.0.0.1:30120/v1/systemone -X POST -H 'content-type: application/json' -d '{
  "model": "jev-latest",
  "state": "Refunds require a receipt. The customer has no receipt.",
  "questions": {"decision": {"type": "noul",
    "instructions": "Is a refund allowed under the rule?",
    "criteria": {"true": "The requirement is met.",
                 "false": "A requirement is missing."}}}
}'

usage.output_tokens is 0. The server applies the temperatures shown above unless a request supplies options.temperature. Each probability vector is normalized over exactly the caller's labels.

Prompt wording and temperatures

Both wordings were evaluated through the same served endpoint. The recommended wording follows the preset rule: use native unless served leads by at least 1.0 percentage point across the 10 suites and an offline check agrees. Temperatures were fitted separately for each wording and answer type on three calibration suites only. No JevBench item was used for these decisions.

Wording	Default temperature	Choice	Yes/no	Ordinal	Mean accuracy, 10 suites
native (recommended)	0.90	0.90	1.15	1.20	69.91 %
served	1.00	1.00	0.90	0.80	70.21 %

For the alternate wording, use --prompt-wording served --default-temperature 1.00 --temperature-by-type choice=1.00,noul=0.90,score=0.80.

Benchmarks

Served endpoint, v2.1

Merged BF16 weights through SGLang 0.5.20 and jev-adapter, one option order, native wording, no system prompt. Accuracy is independent of the fitted temperature.

Suite	Questions	Accuracy
Judge proxy	600	87.50 %
Hard proxy	600	44.17 %
Stated-distribution probability	1,036	78.38 %
Realistic transfer	600	90.17 %
MuSiQue multiple choice	500	69.80 %
SQuAD 2.0 unanswerable	500	79.80 %
ContractNLI	500	62.20 %
PAWS-X English	500	74.40 %
Hard decisions	800	51.50 %
Consistency	1,200	61.17 %

The following public tiers are reported separately and were excluded from wording, temperature, checkpoint and quantization decisions. These are our measurements, not official sealed-set scores.

Suite	Questions	Accuracy
JevBench public easy	48	100.00 %
JevBench public standard	72	86.11 %
JevBench public hard	111	47.75 %

At the fitted temperatures, public-hard 10-bin ECE is 0.212 and mean total-variation distance on the stated-distribution validation set is 0.124.

Offline six-suite result, v2.1

This separate transformers run used native wording, no system prompt and T=1. It is not the served endpoint result. The three public JevBench tiers below are report-only and were excluded from wording, temperature, checkpoint and quantization decisions.

Suite	Questions	Accuracy
JevBench public easy	48	100.00 %
JevBench public standard	72	93.06 %
JevBench public hard	111	47.75 %
Judge proxy	600	88.00 %
Realistic transfer	600	89.33 %
Stated-distribution probability	1,036	77.99 %

Public classification and decision suites, v2.1

Served endpoint, 400 cases per suite, seed 13: AG News 84.5 %, typed decisions 67.3 %, nine-language MASSIVE intent mean 83.4 %, email spam 92.0 %, phishing 88.8 %. Results vary by language and task.

The full report in docs/BENCHMARKS.md and the classification tables in docs/public-classification-suites.md are historical; use the v2.1 results above for this checkpoint. The historical v2 benchmark chart shows earlier-version measurements.

Model details

• Base model: mistralai/Ministral-3-3B-Instruct-2512-BF16, revision b6d637bef2393152b3da2b2fde72eecdee30557e (Apache-2.0).
• Adapter architecture: LoRA r=16, α=32, dropout 0 on q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj and down_proj of the language model.
• Adapter file: adapter_model.safetensors, 135,113,048 bytes, SHA-256 e903d55404c9a2311b7188ca2f71e7dc6333247c20662606168dc4b61075d93b.
• Merge: 182 language-model tensors changed, none outside the language-model projections; maximum absolute weight change 0.0023.
• Serving: native wording, no system prompt, temperatures (choice 0.90, yes/no 1.15, ordinal 1.20), SGLang 0.5.20 and jev-adapter; 8,192-token context, one caller-supplied option order and no rotation ensemble.

Training data

The public source inventory shared with Standard One 8B is listed below. Training data is synthetic and format-augmented decision data plus decision items converted from public datasets (listed below); the JevBench public tiers used only for evaluation carry MIT. Full per-cohort breakdown: docs/BENCHMARKS.md.

Public datasets used (train splits where the dataset has one; licence as stated by each dataset; labels come from the datasets, distractor options are generated by code):

Dataset	Licence
SQuAD 2.0	CC BY-SA 4.0
ARC	CC BY-SA 4.0
BoolQ	CC BY-SA 3.0
CommonsenseQA	MIT
HellaSwag	MIT
Banking77	CC BY 4.0
Bias in Bios	MIT
Bitext customer support	CDLA-Sharing-1.0
CLINC150	CC BY 3.0
Amazon Counterfactual	CC BY 4.0
DBpedia-14	CC BY-SA 3.0
Dolly 15k	CC BY-SA 3.0
GoEmotions	Apache-2.0
MASSIVE	CC BY 4.0
Twitter Financial News Sentiment	MIT
HelpSteer3	CC BY 4.0
HelpSteer2	CC BY 4.0
2WikiMultihopQA	Apache-2.0
HotpotQA	CC BY-SA 4.0
MuSiQue	CC BY 4.0
QASC	CC BY 4.0
DROP	CC BY-SA 4.0
GSM8K	MIT
TempReason	CC BY-SA 3.0
MultiNLI	OANC / CC BY-SA 3.0 / CC BY 3.0
PAWS	Google terms, free for any purpose
PAWS-X	Google terms, free for any purpose
SNLI	CC BY-SA 4.0
WANLI	CC BY 4.0
ContractNLI	CC BY 4.0
CUAD	CC BY 4.0
ShARC	CC BY-SA 3.0
Jailbreak classification	Apache-2.0
Prompt injections	Apache-2.0
Aegis AI Content Safety 2.0	CC BY 4.0
Jigsaw Toxic Comment Classification (mirror of the Kaggle data)	CC0 (data); comment text CC BY-SA 3.0 (Wikipedia)
Measuring Hate Speech	CC BY 4.0
Image safety classes	MIT

Historical upstream-id mapping: docs/BENCHMARKS.md.

Limitations

• Results depend on the task distribution and the serving path. Refit temperatures when using a materially different distribution.
• At most 26 options per question (one uppercase letter per option, A–Z).
• Served and offline probabilities can differ on identical prompts. Use the served results to estimate endpoint behavior.
• Korean is a small share of the multilingual training mixture.
• The card reports text benchmarks; it does not establish accuracy on image inputs.
• The sealed JevBench set has not been measured for this version.

Licence

Adapter weights, merge recipe and this card: Apache-2.0. Base model mistralai/Ministral-3-3B-Instruct-2512 (and -BF16): Apache-2.0 per its Hugging Face model card, which adds that the model must not be used in a way that infringes, misappropriates, or otherwise violates any third party's rights. jev-adapter and SGLang: Apache-2.0. The JevBench harness and public tiers used for evaluation: MIT; other benchmark items keep their own upstream terms.

Citation

StandardThinking/StandardOne-3B-LoRA (this repository, adapter + merge recipe) · StandardThinking/StandardOne-3B (merged weights) · StandardThinking/StandardOne-8B (server code).