blockboiarb/indigo-violet-1

Indigo Violet 1

Indigo Violet answers typed questions about a text with probabilities: yes/no questions, multiple choice and ordered scores. Every question of a request is answered from one forward pass of a 0.6B-parameter model cut to 18 of its 28 layers, with no text generation.

Violet is the smallest Indigo model. The family is named after purple flowers, from the smallest plant to the largest: Violet, Iris, Lilac, Wisteria.

• Code: https://github.com/rwondo/indigo-violet-1 (the indigo-decisions Python package, import name indigo)
• Base model: Qwen/Qwen3-0.6B at revision c1899de289a04d12100db370d81485cdf75e47ca
• Release: indigo-violet-1, served in float16 on GPU and float32 on CPU

Usage

Install the package from its repository (Python 3.10 or newer, PyTorch 2.4 or newer):

pip install "git+https://github.com/rwondo/indigo-violet-1"

import indigo

model = indigo.load("blockboiarb/indigo-violet-1")
model.decide(
    "Order 1182 arrived two weeks late and the customer asks for a refund.",
    {
        "refund": {"type": "noul", "instructions": "Does the customer ask for a refund?"},
        "team": {
            "type": "choice",
            "instructions": "Which team should handle this?",
            "criteria": {"billing": "Refunds and invoices", "shipping": "Deliveries"},
        },
        "urgency": {
            "type": "score",
            "instructions": "How urgent is this?",
            "criteria": ["Not urgent", "Somewhat urgent", "Very urgent"],
        },
    },
)
# {"refund": {"type": "noul", "noul": P(yes)},
#  "team": {"type": "choice", "probabilities": {"billing": p, "shipping": p}},
#  "urgency": {"type": "score", "probabilities": {"0": p, "1": p, "2": p}}}

indigo.load checks the sha256 of the weight files listed in indigo.json, downloads the base model at its pinned revision (about 1.5 GB), merges the adapter, keeps the first 18 layers and installs the readout head.

Results

Kaggle T4 GPU, float16, this release's weights and calibration, one request at a time.

Benchmark	Indigo Violet 1
JevBench-mini accuracy / Brier (1,938 questions with gold answers)	0.807 / 0.339
JevBench-mini accuracy by type: yes/no, choice, score	0.876, 0.852, 0.370
jabr classifier-benchmark, micro accuracy (944 cases)	0.737
jabr accuracy by type: choice, yes/no, score	0.837, 0.659, 0.682
JevBench-mini latency per request (about 10 questions), p50 / p95	49.2 ms / 68.5 ms
jabr latency per question, mean	11.0 ms

• On the same benchmarks Qwen3-4B-Instruct-2507, run zero-shot and uncalibrated with an earlier version of the same harness, scored 0.850 accuracy and 0.268 Brier on JevBench-mini and 0.793 on jabr. Indigo Violet trades some accuracy for speed: 18 layers of a 0.6B model per request.
• JevBench's coherence score (overall 0.880 [0.867, 0.893]) is not a quality measure on its own: answering every question uniformly scores 0.894 on it.

Training

• Distillation. The teacher, Qwen/Qwen3.5-27B (Apache-2.0), labelled every yes/no probe of the training texts with its yes/no logit gap. The student, Qwen/Qwen3-0.6B with a LoRA adapter (rank 16, alpha 32; down_proj, gate_proj, k_proj, o_proj, q_proj, up_proj, v_proj projections), was trained to match the teacher's gaps (rescaled by a fitted linear map) plus a loss on the datasets' own gold labels. The served model keeps the first 18 layers and reads the gap from a readout head trained on layer 18.
• Data. A training set of 41,501 texts from 27 sources (26 datasets; LEX GLUE gives two). All come from train splits except LegalBench, whose texts come mostly from its test splits (its train splits hold a few examples per task). No text comes from JevBench, jabr, their source datasets or their generators, and the texts were screened against those benchmarks for exact and near-duplicates.
• Held out. 2,981 texts were held out to select the checkpoint: a random 4%, every HotpotQA text and a few question families. The model was trained on the other 38,520 texts.

Dataset	Licence	Texts in the training set
ShARC	CC BY-SA 3.0	4,628
stasvinokur/cve-and-cwe-dataset-1999-2025	CC0 1.0	3,000
legacy-datasets/banking77	CC BY 4.0	2,490
clinc/clinc_oos	CC BY 3.0	2,490
DFKI-SLT/few-nerd	CC BY-SA 4.0	2,000
nvidia/HelpSteer2	CC BY 4.0	2,000
coastalcph/lex_glue (LEDGAR)	CC BY 4.0	1,997
ucberkeley-dlab/measuring-hate-speech	CC BY 4.0	1,750
DoDataThings/us-bank-transaction-categories-v2	MIT	1,500
google/boolq	CC BY-SA 3.0	1,500
tasksource/esci	Apache-2.0	1,500
FuzzyLabs/ntsb-accident-extraction	CC0 1.0	1,500
allenai/prosocial-dialog	CC BY 4.0	1,500
wenhu/tab_fact	CC BY 4.0	1,500
OpenAssistant/oasst2	Apache-2.0	1,499
nguha/legalbench	CC BY 4.0, CC BY-SA 4.0 or MIT, per task	1,479
nvidia/Aegis-AI-Content-Safety-Dataset-2.0	CC BY 4.0	1,246
fancyzhx/dbpedia_14	CC BY-SA 3.0	1,000
Anthropic/hh-rlhf (red-team attempts)	MIT	1,000
NOAA NCEI Storm Events Database (2015-2024)	US federal work, public domain	1,000
coastalcph/lex_glue (UNFAIR-ToS)	CC BY 4.0	1,000
nvidia/Nemotron-PII (synthetic)	CC BY 4.0	800
nvidia/HelpSteer3	CC BY 4.0	750
hotpotqa/hotpot_qa	CC BY-SA 4.0	750 (held out, not trained on)
Zhengping/UNLI	MIT (SNLI text: CC BY-SA 4.0)	750
lasha-nlp/CONDAQA	Apache-2.0	474
google/civil_comments	CC0 1.0	398

Some datasets are share-alike (CC BY-SA): ShARC, Few-NERD, BoolQ, DBpedia-14, HotpotQA, the SNLI text in UNLI and some LegalBench tasks. Whether share-alike terms reach weights trained on such data is unsettled; check them for your use. The HelpSteer prompts come mostly from ShareGPT conversations, and ProsocialDialog was drafted with a GPT-3 model.

How it works

1. The LoRA adapter is merged into Qwen/Qwen3-0.6B at revision c1899de289a04d12100db370d81485cdf75e47ca.
2. Only the first 18 decoder layers are kept. The final norm then applies to layer 18, and the readout head reads the gap z_yes - z_no = w . final_norm(h_18) + c at the last token of a yes/no prompt.
3. Each question becomes yes/no probes. A yes/no ("noul") question is one probe; instructions that do not end with "?" are asked as Is the following statement true? {statement}. A choice or score question has one probe per option or level: {instructions}\nCandidate answer: {option}\nIs the candidate answer correct?.
4. All probes of a request run in one forward pass with a tree attention mask: each probe sees the text and its own question only, so the result is the same as running every probe alone (up to floating-point rounding).
5. Calibration turns gaps into probabilities: P(yes) = sigmoid(8 * gap) for a noul, softmax(gap / 0.125) over the options of a choice and softmax(gap / 0.125) over the levels of a score.

The adapter holds LoRA weights for all 28 layers, as trained. Only layers 0 to 17 are served: the weights of layers 18 to 27 are merged and then dropped with those layers, so they have no effect.

Limits

• English. The training texts are almost all English.
• Probabilities. Gaps become probabilities through a fixed sharpening, not a fitted calibration (JevBench-mini Brier 0.339).
• Short texts. Training texts were at most 1,024 tokens. Longer texts run (nothing is truncated) but are outside what the model was trained on.
• Unfamiliar tasks. Accuracy drops on task types unlike the training data, and knowledge-heavy questions (exams, maths, science) are beyond a model this size.
• No truncation. A request whose forward pass (instruction header, text and every probe) is longer than max_tokens is refused with TooLongError before anything runs. The default is 8192 tokens, and max_tokens can be raised up to the base model's context length of 40960 tokens.
• Saturation: with the choice temperature of 0.125, an option that leads the others by more than about 4.6 in gap gets a probability of exactly 1.0 in float64. For a ranking across requests, use the raw gaps (model.gaps(...)).

Decision Index

With the decision-index extra installed, the package runs as a Decision Index engine:

pip install "indigo-decisions[decision-index] @ git+https://github.com/rwondo/indigo-violet-1"
python -m decision_index run --engine indigo.decision_index:IndigoEngine --option model=blockboiarb/indigo-violet-1

The engine renders every request the same way, refuses passes over max_tokens as unsupported and gives every option a probability.

Files

File	Content	sha256
adapter_model.safetensors	LoRA weights (PEFT format)	33e18bbe133b19e180b7ce3427e4c78f36544ea2ecad2d441349bb1e7d050da9
adapter_config.json	LoRA configuration	7a19981627759a0200c80cd3a78891c11d51a597b177e760028821822d471a80
readout.safetensors	readout head: w [1024] and c [1], float32	53063ede4feae2a52bbb158dcce4b8ee5a6df07a0f304ff91d1d0842aadbb51b
indigo.json	base model and revision, depth, prompt, calibration, engine defaults, file digests

License

The files in this repository are released under the Apache License 2.0 (see LICENSE). The base model, Qwen/Qwen3-0.6B, is also released under the Apache License 2.0. The training data keeps its own licences (table above).