basal-omni-1.1-max (experimental task model)

basal-omni-1.1-max (12B) is basal-omni-1.0-max with one short round of extra training on three of the task formats of basal 2.1: verify, compress and extract (classify was not trained; its results carry over). It is better at the tasks and slightly behind 1.0-max on general decisions, so basal-omni-1.0-max stays the default model; use 1.1-max when the tasks are your main use, and check it on your own data.

Like every basal model it does not write text: it reads a state (a message, a document, retrieved passages, an agent trace) and answers typed questions about it with a calibrated probability for each allowed answer, in one forward pass. The tasks are built from those decisions by the engine (see Tasks).

Results

Both models served with the basal engine on one H100 (--mode eager, bf16, one option order), each with its own calibration file, on the same items with the same clients.

Tasks

Measured with basal-serve 2.1:

• What it saw in training: the train splits of VitaminC and QASPER, and synthetic documents for extract from a different generator than the test invoices. CLINC150, BANKING77 and RAGTruth were not used in training, so those three rows are results on unseen datasets.
• Overlap check: 27 test items (12 VitaminC, 15 QASPER) share a 13-word span with training rows. Without them the gains are +4.2 points on VitaminC (paired 95% interval +2.1 to +6.2) and +2.7 on QASPER (+0.8 to +4.9); the mean of the six sets is still 85.1.
• Whole documents: all 12 invoice fields correct on 60.5% of invoices (1.0-max 52.5%); on the unseen layouts 38.8% of documents fully correct (1.0-max 22.8%).
• Automation: automated is certified for classify at a 5% target error: on a held-out labelled split, 38.5% of answers were automated, and 2.2% of those were wrong. No threshold could be certified for classify at 1% or for verify; there automated is null: review those answers, or fit task_thresholds on your own labelled data.

General decisions

• Werdykt v1: 0.821 (95% interval 0.811–0.830; Polish 0.819, English 0.822), −0.2 points against 1.0-max, inside the interval. Served with --deterministic; on this model the default mode gave the same answer on 2,000 of 2,000 items. Its probabilities are slightly better calibrated (ECE 0.051 vs 0.054), but it automates fewer decisions at the 1% target: 57.7% at 1.23% observed error (1.0-max 60.1% at 1.45%).
• ContractNLI: the largest loss, 0.583 against 0.720 (−13.7 points, paired 95% interval −19.3 to −7.7). On plain entailment choices 1.1-max too often picks "does not mention enough" (189 of 300 items, 1.0-max 117): the verify training made "not stated" more likely. For entailment questions outside verify, use 1.0-max.

Accuracy by document length

Accuracy of basal-omni-1.1-max on Werdykt (5,000 items, Polish and English; a hidden set, so only these totals are shown) by the length of the state, in characters (a token is about 2 characters; 1,000 characters is about half a page). 82.1% overall. 95% Wilson intervals in brackets.

Short states are mostly rules and abstention items, which are harder for every basal model, so read the rows as accuracy per bucket, not as the effect of length alone. The long-document family keeps the task and varies only the length:

Run 2026-10-15 with basal-serve --mode eager on one NVIDIA H100 80GB HBM3 (bf16). The intervals are about the items, not about run-to-run noise.

By prompt length in tokens (held-out tests)

The held-out tests of the Results table by the length of the whole prompt as served (state, question and options; this model's tokenizer), with the Results table's predictor. Accuracy (items):

Quick start

1. Install into a fresh uv environment:

uv venv --python 3.12 --allow-existing ~/basal-env && source ~/basal-env/bin/activate
uv pip install torch==2.11.0 torchvision==0.26.0 --index-url https://download.pytorch.org/whl/cu128
uv pip install "basal[media] @ https://github.com/rkinas/basal/archive/refs/tags/v2.5.3.tar.gz"

2. Start the server. It prints basal: model ... ready:

basal-serve --model Remek/basal-omni-1.1-max --mode eager --media on --port 8000

The eager mode (PyTorch, bf16, about 24 GB of GPU memory), as for 1.0-max.

3. Ask questions (choice, noul, score, multi, act, tool and the tasks below), exactly as with 1.0-max. Full API: API.

Faster: --mode vllm. With vLLM 0.31.0 (its PyPI build needs NVIDIA driver 580 or newer), in its own environment:

uv venv --python 3.12 --allow-existing ~/basal-vllm && source ~/basal-vllm/bin/activate
uv pip install vllm==0.31.0 "basal @ https://github.com/rkinas/basal/archive/refs/tags/v2.5.3.tar.gz"
basal-serve --model Remek/basal-omni-1.1-max --mode vllm --max-len 16384 --gpu-memory 0.9 --port 8000

• On one H100 (basal-loadtest, 16 requests in flight, latency p50 / p95): one question per request 16.7 decisions/s, 65 / 93 ms (eager: 4.77, 183 / 348 ms); 5 questions over one 3,000–10,000-token state 10.8 questions/s, 0.50 / 0.90 s per request (eager: 0.72, 8.8 / 14.1 s).
• It uses CALIBRATION.vllm.json, fitted for this path. Against --mode eager it gives the same answer on 98.9% of our dev decisions (99.97% where eager's top answer leads by 0.2 or more), accuracy +0.1 point. It automates 47.0% of decisions at the 1% target (eager 47.1%) and 82.7% at the 5% target (eager 86.0%; the vLLM file's 5% thresholds are stricter).
• Images and audio ("media", with --media on) need basal 2.5 or later; before 2.5 this path is text only.
• One option order, and no --deterministic: vLLM batches requests its own way, so a decision near a tie can change with the load. One request at a time, basal 2.1 returns the same probabilities as 2.0, bit for bit; with 32 requests in flight, vLLM's batching changes about 1% of top answers between any two runs. The tasks were measured in --mode eager.

Apple Silicon, text only: the 8-bit port basal-omni-1.1-max-MLX runs on --mode mlx (see its card).

Tasks

basal 2.1 adds four task types on the same POST /v1/basal API (experimental: the API may change in a minor release):

• classify: labels with a rubric each.
• verify: statements about the text; each one is supported, contradicted or not stated.
• compress: keeps the sentences, lines or paragraphs that a question needs.
• extract: fills the fields of a JSON schema, copied exactly from the text with character offsets.

Each task is expanded into the model's own decisions over one shared input, so every answer comes with calibrated probabilities. Details: TASKS.md.

curl -s localhost:8000/v1/basal -H 'content-type: application/json' -d '{
  "state": "I was charged twice for my card payment yesterday.",
  "questions": {"intent": {"type": "classify", "none": true, "labels": [
    {"name": "card_payment_duplicate", "rubric": "a card payment was taken more than once"},
    {"name": "card_lost", "rubric": "the card is lost or stolen"},
    {"name": "refund_status", "rubric": "asks about a refund in progress"}]}}}'

The answer (answers.intent, rounded to 4 places):

{"type": "classify", "label": "card_payment_duplicate", "top_label": "card_payment_duplicate",
 "scores": {"card_payment_duplicate": 0.9995, "card_lost": 0.0002, "refund_status": 0.0000, "none": 0.0003},
 "confidence": 0.9995, "automated": {"0.01": null, "0.05": true}, "experimental": true}

"0.05": true: the confidence is above the threshold certified for classify at a 5% target error, so the answer may be used without review at that rate.

Images and audio

The same as basal-omni-1.0-max: with --media on, images and audio go in "media" and every question type reads them, zero-shot (the extra training was text only). Media answers were not re-measured for 1.1-max; extract does not read text fields from images yet.

Speed

The same architecture as 1.0-max. On Werdykt's 5,000 items (16 concurrent clients, one H100, --mode eager): 4.12 decisions/s, 86 ms p50 / 1.40 s p95 (1.0-max: 4.17, 76 ms / 1.36 s). classify is one decision; the other tasks cost one decision per claim (verify), unit (compress) or field (extract, in two packed rounds), all over one shared state.

Files

• model.safetensors, tokenizer, chat_template.jinja, processor_config.json: the model (Gemma 4 architecture, 12B, 262k vocabulary; the image and audio encoders are included and unchanged).
• CALIBRATION.json: per-type temperatures and certified confidence thresholds for 1% / 5% error, fitted on the served engine (--mode eager, bf16, one option order), and task_thresholds (classify at 5%).
• CALIBRATION.o2.json: the same for --orders 2, without task thresholds (automated is null for tasks there).
• CALIBRATION.vllm.json: the same for --mode vllm (one option order), without task thresholds.
• basal.json: prompt format 2 and the readout protocol.
• SHA256SUMS: checksums of every file.

Training

• Start: basal-omni-1.0-max (merged weights).
• Method: LoRA (rank 32) on the language-model layers, learning rate 2e-5, one epoch of 84,998 decisions (1,325 steps of 64), merged into the weights. The vision and audio parts were not trained.
• Data: 39,998 task decisions in the engine's own task prompts (extract line and span choices on synthetic documents 28,000; claim verification from the VitaminC train split 7,945; paragraph relevance from the QASPER train split 4,053) and 45,000 decisions replayed from the same data as 1.0-max, against forgetting.
• Calibration: refitted for this model on the same calibration data and with the same checks as 1.0-max (per-type temperatures and certified thresholds); task thresholds fitted on separate labelled task splits (70% fit / 30% certify).

Limitations

• Experimental, and not the default: for general decisions basal-omni-1.0-max is slightly better (see Results).
• Entailment outside verify: ContractNLI −13.7 points; plain entailment choices lean to "not stated".
• Automation: fewer decisions automated at the 1% target than 1.0-max on Werdykt (57.7% vs 60.1%); for tasks, only classify at 5% has a certified threshold.
• extract: currency is the weakest invoice field (62.5%); no arrays of strings or objects (line items) yet; no text fields read from images.
• Exact arithmetic and thresholds are a weak spot of one-pass decisions: compute them in code.
• Review: decisions with serious consequences for people should be reviewed by a person.

Citation

@software{kinas2026basal21,
  title   = {basal-2.1: typed-decision models and inference engine for Polish and English},
  author  = {Kinas, Remigiusz},
  year    = {2026},
  version = {2.1.0},
  url     = {https://github.com/rkinas/basal}
}

@techreport{kinas2026basal,
  title       = {basal-1.0: Reliable, Highly Optimized Typed Decisions for Polish},
  author      = {Kinas, Remigiusz},
  institution = {ai5},
  year        = {2026},
  type        = {Technical report},
  doi         = {10.5281/zenodo.23022986},
  url         = {https://doi.org/10.5281/zenodo.23022986}
}

License and attribution

Apache-2.0. Fine-tuned from Remek/basal-omni-1.0-max, itself fine-tuned from google/gemma-4-12B-it (Apache-2.0).