h t t p s : / / h u g g i n g f a c e . c o / R e m e k / b a s a l - 2 . 0 - m a x

Ingest: Basal 2.0 max model card

Type: types/ingest-report.md

Classification

A producer's release card documenting a decision model, its serving interface, benchmarks, deployment variants and limitations. It is not an independent evaluation or a scientific paper.

Author: Remigiusz Kinas, the named producer and software citation author. This gives direct access to the intended design but an interested source for performance claims.

Summary

Basal 2.0 max is an 11B Polish/English typed-decision model fine-tuned from Bielik-PL-11B-v3.0-Instruct through earlier Basal checkpoints. Its documented engine scores caller-supplied options using next-token option-code logits in one forward pass, then applies fitted temperatures; it does not generate prose. Supported questions include classification, yes/no checks, scores, multiple labels, actions and tool selection without arguments. The card reports Werdykt macro accuracy of 0.784 versus 0.773 for Basal 1.5 and lower calibration error, but worse results on some other sets. These are release-system comparisons within a finite-option decision setup, with changed prompts, calibration and serving configurations, not an isolated training effect or comparison against alternative task decompositions. Held-out calibration reports 41.9% automation at 2.51% observed error for the 5% threshold; the shipped 1% threshold defers everything. The card requires local validation, assigns exact arithmetic to code, and reports port agreement rather than universal interchangeability.

Quotes

basal-2.0-max is part of basal-2.0, a family of typed-decision models for Polish and English; it is the most accurate model of the family. Built on Bielik-PL-11B-v3.0-Instruct. Trained further from earlier basal checkpoints. --- kb/sources/.snapshots/basal-2-0-max-20261024.md @ sha256:f6fadefd922e7bc737ad37f85ae473ca60ecf9bfe8202493b32866d8d7ee1b6c

What it is. The model reads a state (a message, a document, retrieved passages, an agent trace, a web page as JSON) and answers a typed question about it (choice, yes/no noul, score, multi, act, tool) by returning a calibrated probability for each allowed answer, in one forward pass, without generating text. The answer never falls outside the options you give, and the probability says how sure the model is. Because the classes are described in the request, one model serves many tasks without retraining: routing, rule and policy checks, questions about long documents, RAG decisions, urgency and risk scores, judging answers, agent and tool decisions, and triage with a confidence threshold. --- kb/sources/.snapshots/basal-2-0-max-20261024.md @ sha256:f6fadefd922e7bc737ad37f85ae473ca60ecf9bfe8202493b32866d8d7ee1b6c

Each model also comes as -FP8 (NVIDIA Hopper and Blackwell), -GGUF (Q8_0 and Q5_K_M; Ollama, llama.cpp) and -MLX-8bit (Apple Silicon). --- kb/sources/.snapshots/basal-2-0-max-20261024.md @ sha256:f6fadefd922e7bc737ad37f85ae473ca60ecf9bfe8202493b32866d8d7ee1b6c

basal-2.0-max, held-out
calibration error (ECE) 0.022
worst question family (ECE) 0.098 (agent)
automated at the 5% error threshold 41.9% at 2.51% observed error
1% error threshold, shipped file defers every decision (no 1% threshold certified on the full fit)

On Werdykt (hidden, measured once, not used to choose the model) the calibration error is 0.044 (basal-1.5: 0.074). The calibration is a file, separate from the weights (CALIBRATION.json, one option order; CALIBRATION.o2.json, two orders), so you can swap it or refit it on your own data with basal-calibrate. Fitted with engine v2.0.0 (commit adc4c71). --- kb/sources/.snapshots/basal-2-0-max-20261024.md @ sha256:f6fadefd922e7bc737ad37f85ae473ca60ecf9bfe8202493b32866d8d7ee1b6c

The model receives a fixed chat prompt with the state, the question and coded options; the decision is the softmax over the next-token logits of the option codes only (one forward pass, fp32 code logits, no text generation), then the per-type calibrated temperature from CALIBRATION.json. Confidence thresholds were chosen on calibration data, before testing, for 1% or 5% error among accepted decisions. --- kb/sources/.snapshots/basal-2-0-max-20261024.md @ sha256:f6fadefd922e7bc737ad37f85ae473ca60ecf9bfe8202493b32866d8d7ee1b6c

  • Validate on your own documents before relying on it: hidden and held-out sets do not cover every domain.
  • Polish world knowledge of a small model is limited: put the relevant facts in the state. Legal rules change.
  • Exact arithmetic and thresholds are a weak spot of one-pass decisions: compute them in code (or "facts": "auto" ).
  • The calibration is cautious on unseen tasks: no 1% threshold is certified in the shipped files, so "max_error": 0.01 on anact ortool question defers every decision; use 0.05, or certify thresholds on your own data withbasal-calibrate .
  • act andmulti over many labels are experimental. Up to 255 options per question (Ollama: 20); the context is 32,768 tokens. The 2.0 models ship no evidence head.
  • Decisions with serious consequences for people should be reviewed by a person. --- kb/sources/.snapshots/basal-2-0-max-20261024.md @ sha256:f6fadefd922e7bc737ad37f85ae473ca60ecf9bfe8202493b32866d8d7ee1b6c

For every question basal returns a probability for each answer; the confidence is the probability of the answer it picks. --- kb/sources/.snapshots/basal-2-0-max-20261024.md @ sha256:f6fadefd922e7bc737ad37f85ae473ca60ecf9bfe8202493b32866d8d7ee1b6c

Connections Found

For the Jev/Basal investigation, this is a producer-described candidate interface and a source of test conditions, not an adoption warrant. Compared with Generation confidence does not by itself certify soundness, Basal supplies a specifically trained and calibrated decision signal rather than generic chat confidence. Its reported selective error and coverage make abstention assessable, but only within the measured configuration and population. They do not establish soundness on KB decisions.

The historical mem review places Jev in optional requested retrieval ranking alongside a local cross-encoder; provider internals and host benefit remain uninspected there. Basal instead documents a state-plus-question finite-option interface. The settled comparison is component placement: retrieval relevance scoring versus caller-defined decision scoring. Neither source establishes shared weights, calibration, interchangeable APIs or relative quality. A retrieval rank is not a truth judgment about a retained fact.

The release comparison also illustrates the boundary in Learning inside a fixed decomposition inherits its mistakes: better scores within supplied states and options do not establish that those states preserve every necessary distinction or that the options cover a KB task. The card does not compare alternative category schemes.

Learning Claims (our opinion)

The documented adaptation consists of merged LoRA fine-tuning and separate calibration: temperatures are fitted by question type and option-count bucket on question families excluded from training, and thresholds are chosen before testing. Calibration files can be replaced or refitted without replacing weights. At inference, new class descriptions condition a decision without retraining; that is task conditioning, not evidence of learning from the request. No online update from accepted or deferred decisions is described.

Against the four conditions of a theory builder, the deployed component has localized textual states, questions and option descriptions, but no identified units stating what its learned weights say. Consumption of those descriptions is part of the documented interface, not a locally tested behavioral result. Content-directed criticism is not demonstrated: benchmark errors and calibration residuals do not by themselves constitute stated criticism of an identified theory. Iteration is documented for continued weight training and fitting, but not for retaining such criticism to revise stated theories. Weights and calibration files persist across requests; request-driven revision does not follow. The card therefore does not establish the component as a theory builder, while leaving it available as one component of a larger builder.

Improved capacity for future action is a separate empirical question. The producer's mixed release results support task-bounded improvement, not uniform progress, and do not isolate fine-tuning from prompt and calibration changes. Training changes weights and calibration changes score processing; neither validates an operator's fixed category set. This source adds a concrete separation between persistent weight adaptation, replaceable probability calibration and request-local task specification. It gives no reason to revise Commonplace's distinction between consuming a stated theory and revising it.

Extractable Value

  1. A bounded KB candidate, not a writing model. Document routing or review-priority suggestions over explicit options are plausible non-mutating trials. Basal cannot author notes, supply missing source facts, generate tool arguments or furnish an evidence head in this release. Classification suitability remains untested on Commonplace material. [experiment]
  2. Evaluate selective coverage alongside error. The reported held-out 5% operating point accepts 41.9% at 2.51% observed error, while the shipped 1% setting accepts nothing. These results concern the card's option-code readout, fitted calibration and evaluation distribution; they do not certify KB error rates or validate the task categories. They identify what a local evaluation must measure rather than a threshold to copy. [quick-win]
  3. Preserve the complete decision configuration. Visible option keys make calibration unvalidated according to the card; one versus two option orders use different files. Type, option count, key visibility, model port and calibration identity belong in any future test record. The distinction is more useful than recording a model name alone. [quick-win]
  4. Keep Jev comparison at the operational boundary. Existing local evidence supports a Jev retrieval-ranking use, not equivalence with Basal. Compare the required input, score interpretation and downstream decision before making a model-quality comparison. [just-a-reference]

Limitations (our opinion)

All performance and calibration results here are producer-reported. No implementation was inspected and no code, model or benchmark was executed. The assertion that calibration holds on new tasks exceeds what family-held-out tests can establish for arbitrary KB distributions. The card itself reports worst-family ECE of 0.098 for agent questions and requires validation on the operator's documents. Neither ECE nor a nominal error target guarantees an individual decision or a new domain's accepted-case error rate.

The release comparison changes more than weights. Prompt format expands the option limit from 10 to 255 and changes the default from two option orders to one; calibration is release-specific. Throughput depends on engine, attention implementation, hardware and load. The reported 16.2 versus 10.2 decisions/s compares the releases' public-board serving setups, not an architecture-only contrast or a general speed advantage over all Basal 1.5 deployments. Compact batching and template-cache ablations concern those engine changes only.

Some reported sets selected training runs, and PL-Real source-document overlap makes the 158-item document-clean subset the primary comparison. Port agreement is measured on development data; agreement with bf16 is not correctness on new material. English-panel mean accuracy declines from 0.7365 to 0.7185, so the release's broad accuracy improvement is not uniform. Many-label action and multi-label decisions remain experimental; the context limit is 32,768 tokens, and Ollama supports fewer options. Arithmetic, changed legal facts and consequential decisions require safeguards outside model confidence. The captured card supplies no direct Jev comparison, local KB cost measurement or end-to-end workflow benefit.

Draft a labelled, non-mutating KB document-routing evaluation dataset with explicit route definitions, ambiguous and out-of-scope cases, and accepted-error/coverage scoring, as preparation for a later separately authorized Jev/Basal investigation rather than model selection or deployment.