StandardThinking/StandardOne-3B-LoRA

Standard One 3B (LoRA adapter)

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.

For ready-to-serve merged BF16 weights, use StandardOne-3B. The server code is in StandardOne-8B.

If you need	Repository
3B adapter weights and merge recipe	StandardOne-3B-LoRA (this repository)
Merged 3B checkpoint	StandardOne-3B
8B adapter weights and merge recipe	StandardOne-8B-LoRA
Larger merged checkpoint and server code	StandardOne-8B

In the reported served evaluations, 3B has a lower median latency on the measured short-request profile; 8B scores higher on the public standard and hard tiers. See Benchmarks for the measurement conditions and limitations.

The figure combines results from different measurement paths. See Benchmarks for served versus offline conditions; measured 24 September 2026.

At a glance

• Send a state and a bounded rubric to receive probabilities for the supplied labels: choice selects among labeled options, noul is yes/no, and score uses an ordinal scale. The endpoint scores the labels in one forward pass without decoding answer text after this adapter is merged and served.
• In the same-run offline comparison with the untuned base, 3B improves on all six suites. The public hard gain is only 1.80 percentage points and should not be read as a broad hard-task gain.
• Probabilities are temperature-scaled and calibration-checked (hard-tier ECE, distribution total-variation) — see Benchmarks below.
• The card metadata lists English and Korean. The training mixture is multilingual, and the reported nine-language MASSIVE intent evaluation has per-language results in docs/public-classification-suites.md; quality should not be assumed uniform across languages. The merged checkpoint retains the Pixtral vision tower and accepts image data URLs; this card does not report a separate image-input benchmark.
• Apache-2.0 throughout: base model, adapter and merged weights.

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)
# then copy the base snapshot's tokenizer / chat template / preprocessor / generation config
# files into ./StandardOne-3B-merged alongside the merged weights.

Then serve ./StandardOne-3B-merged with the sglang.launch_server and jev-adapter setup in StandardOne-8B's QUICKSTART.md, substituting the merged directory for --model-path (served model name standard-one-3b, no system prompt, --default-temperature 1.45). That guide also has the virtual-environment installation, request format and curl example.

Benchmarks

Served endpoint results (the release configuration). Merged BF16 weights through SGLang 0.5.20 and jev-adapter, native wording, no system prompt, one option order, T=1.45. Jev 1.13 was measured on the same items through its hosted endpoint; its probabilities are raw, with no temperature applied. These are our measurements, not official sealed-set JevBench scores.

Suite	Standard One 3B	Jev 1.13
JevBench public easy (48)	100.00 %	100.00 %
JevBench public standard (72)	86.11 %	98.61 %
JevBench public hard (111)	48.65 %	72.07 %
judge proxy (600: routing + answer adequacy)	85.67 %	90.50 %
realistic transfer set (600)	88.17 %	86.67 %
stated-distribution probability (1,036)	80.12 %	72.97 %
hard proxy (600)	43.67 %	54.83 %

Offline comparison with the untuned base. This separate transformers runner used native wording, no system prompt, one option order, T=1. The base and tuned checkpoint were scored by the same offline path; these numbers are indicative of the base-model change, not the served scores above.

Suite	Untuned base	Standard One 3B
JevBench public easy (48)	97.92 %	100.00 %
JevBench public standard (72)	70.83 %	91.67 %
JevBench public hard (111)	47.75 %	49.55 %
judge proxy (600: routing + answer adequacy)	63.17 %	85.67 %
realistic transfer set (600)	62.50 %	88.67 %
stated-distribution probability (1,036)	31.56 %	79.73 %

Against the untuned base, all six suites improve in this offline run; the public hard gain is only 1.80 percentage points. Served and offline probabilities differ even on identical prompts, so use the served table for expected endpoint behavior. Hard-tier ECE at served T is 0.158 (T=1.45) against Jev 1.13's 0.099 raw, and mean TV to the stated distributions is 0.135 at served T against Jev's 0.192 raw; full calibration table: docs/BENCHMARKS.md.

Speed — raw serial latency on one H200 with SGLang 0.5.20, using a 242-decision profile averaging about 280 input tokens per decision. The 22.6 ms figure is p50 for this profile, not a latency guarantee for other request lengths, concurrency or hardware. Qwen checkpoints are untuned and shown for speed only; no accuracy comparison is implied.

Model	p50	p95	Input tokens/decision
Standard One 3B	22.6 ms	33.2 ms	≈280
Standard One 8B	25.8 ms	41.9 ms	≈280
Qwen3-8B (untuned)	28.5 ms	57.2 ms	278
Qwen3.5-4B (untuned)	48.8 ms	72.7 ms	283

Throughput has been measured for the 8B only (see docs/BENCHMARKS.md); no equivalent sweep has been run for the 3B.

On the public classification and decision suites (400 cases/suite, seed 13, served endpoints): AG News 83.8 %, typed decisions 69.3 %, MASSIVE intent mean 84.0 %, email spam 91.8 %, phishing 89.5 %. Full table, per-language and per-workflow breakdown: docs/BENCHMARKS.md and docs/public-classification-suites.md.

A JevBench v1.4.1 run has been requested; the sealed-set result is not yet available.

Full report: docs/BENCHMARKS.md.

Model details

• Base model: mistralai/Ministral-3-3B-Instruct-2512-BF16, revision b6d637bef2393152b3da2b2fde72eecdee30557e (Apache-2.0).
• Adapter: LoRA r=16, α=32, dropout 0, on q_proj k_proj v_proj o_proj gate_proj up_proj down_proj of the language-model projections only (vision tower and multimodal projector excluded), 24,707,072 trainable parameters, PEFT 0.21.0. Adapter file adapter_model.safetensors, 135,113,048 bytes, sha256 ae5fe8e3be2e9186e6504e48ad2fe7eea5537e7e8382a0cca2ca49d95605f7bf.
• Merged BF16 checkpoint (as published in StandardThinking/StandardOne-3B): merging this adapter into the base changes 182 tensors, none outside the language-model projections, maximum absolute weight change 0.0025.
• Serving details: native chat-template wording, no system prompt, fixed temperature (T = 1.45, fitted on held-out calibration data); served model name standard-one-3b behind stock SGLang 0.5.20 via jev-adapter (POST /v1/systemone); single caller-supplied option order, no rotation ensemble; 8,192-token context.

Path	Contents
adapter_model.safetensors, adapter_config.json	The LoRA adapter
merge.py	Loads the base model, applies this adapter, saves the merged BF16 checkpoint
docs/, docs/public-classification-suites.md	Full benchmark report, figures (shared with 8B repo), per-language/per-workflow numbers
SHA256SUMS, release-manifest.json, MERGE_REPORT.json, evidence/, LICENSE, README.md	File hashes, training manifest, merge report, supporting artifacts, licence, this card

The merged BF16 checkpoint is published in StandardThinking/StandardOne-3B; the server code and full quick-start guide live in StandardThinking/StandardOne-8B.

Training data

Trains on the identical mixture as StandardThinking/StandardOne-8B-LoRA — same cohorts, same row counts, same overlap-audit result, not a reduced subset. Training data spans document and field normalisation, judge/routing and answer-adequacy, stated-distribution probability, verification/scoring, adequacy-rubric style, and game- and Tetris board-state cohorts — every row is synthetic and solver-generated (deterministic generators with visible checkers; a subset was teacher-reviewed), licensed Apache-2.0 (ours); the JevBench public tiers used only for evaluation carry MIT. Full per-cohort breakdown: docs/BENCHMARKS.md.

An exact-text overlap audit against the public JevBench tiers found 0 exact scenario matches and 181 exact instruction matches — rows in two adequacy-rubric cohorts whose entire instruction field, a generic 58-character adequacy question, is byte-identical to one public hard-tier instruction (0.05 % of the 359,497-row training mixture). These rows are kept and disclosed here rather than regenerated, since the overlap is limited to one rubric question's wording and never touches a scenario or an answer.

Limitations

• Public hard tier: the served 3B score is 48.65 %, versus 72.07 % for Jev 1.13. In the separate offline base comparison, Standard One 3B scores 49.55 %, only 1.80 percentage points above the untuned base's 47.75 %.
• Served probabilities are temperature-scaled by a single global value; if you apply this model to a materially different question distribution, re-fitting that temperature is advisable rather than assuming 1.45 transfers.
• At most 26 options per question (one uppercase letter per option, A–Z).
• The sealed JevBench set has not been measured for this model.
• Served and offline probabilities can differ on identical prompts (mean total-variation ≈0.07 on the hard tier); served numbers are treated as authoritative.
• Korean is a small share of multilingual training alongside English.
• The card reports text benchmarks; it does not establish decision accuracy on image inputs.
• Ten-way support triage (34–40 %) and RAG passage relevance (58–61 %) are weak zero-shot; fine-tune for those.

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).