NB-ASR Qwen3 Lunde v07 — 100k longform hesitation

This is an experimental Norwegian Bokmål Qwen3-ASR-0.6B checkpoint from the NB-ASR project at the National Library of Norway. It starts from the selected 1,200-step long-context recovery model derived from the completed 100,000-step run, then adds 1,000 optimizer steps of hesitation-focused post-training. It is an evaluation checkpoint, not a production release.

The model supports two transcription styles:

• No text prompt: reading-optimised Norwegian Bokmål.
• <verbatim>: closer reproduction of the spoken form, including canonical hesitation forms ee…, ehm…, mm…, and mhm.

<summary> exists in the bundled custom tokenizer but was not trained as an output style. Summary-prompt evaluation is diagnostic only.

Model lineage

• Architecture: Qwen/Qwen3-ASR-0.6B
• Direct parent: NbAiLab/nb-asr-qwen3-lunde07-mixed-nob-100k-longfix
• Parent revision: 8daa84f92c2e90556d5f83e2baeab46d8f002ee6
• Parent training: 100,000 base steps plus selected 1,200-step long-context recovery
• Published hesitation checkpoint: 1,000 post-training steps
• Training precision: bf16; all model parameters trainable
• Hardware: 4 NVIDIA H200 GPUs
• Peak learning rate: 5e-6, cosine decay
• Gradient checkpointing: non-reentrant

Hesitation boost

The post-training data came from the high-confidence silver dataset NbAiLab/nb-asr-hesitation-hc-silver at revision 003b37ea20b6b6a9f3dfdb0ee897e68a4f30e6e5. Candidate verbatim/reading pairs were filtered for canonical hesitation evidence and checked by the NbAiLab/nb-asr-qwen06b-disfluency-reranker. Recording groups were isolated between train, validation, and test.

Each accepted audio item was expanded into two supervised mappings:

• audio + <verbatim> → verbatim target with canonical hesitation notation;
• audio + empty prompt → reading-optimised target.

Training used 56,971 silver training pairs expanded to 113,942 focused rows, plus 28,486 balanced replay rows. The resulting 142,428-row manifest was approximately 80% focused data and 20% replay, with the two styles exactly balanced. A paired-session sampler prevented both targets for the same audio from occurring in one optimizer update.

Step 1,000 was chosen conservatively from validation and protected regression suites. Canonical hesitation F1 had largely plateaued by 800–1,000 steps; later checkpoints did not provide a material, stable advantage.

These results are from 500 validation recording pairs evaluated with both prompts. The recording-group-isolated test set was deliberately not used to select step 1,000 and was not subsequently evaluated at this checkpoint.

Long-form evaluation

The direct parent repaired severe long-input omissions using a 1,200-step recovery phase with a 50/50 optimizer-step mixture of ordinary short ASR and synthetic long examples around 300, 600, 1,200, and 1,700 seconds. Its intended support target was approximately five minutes.

The hesitation boost retained short, streaming, and one natural near-five-minute cohort, but did not preserve all of the parent's synthetic/extended long-form reliability:

The model had one generation failure in the 309–319-second cohort and one in the 1,800-second stress set. Therefore, this checkpoint should be treated as hesitation-optimised; robust five-minute support is not established despite the longform lineage in its name.

Inference

Use the processor, tokenizer, and chat_template.jinja bundled with this repository. Audio must be mono 16 kHz floating-point samples. The key style contract is:

# Empty text means reading-optimised output.
style_prompt = ""

# Use this instead for verbatim output with canonical hesitations.
style_prompt = "<verbatim>"

The raw assistant prefix used in training is language Norwegian<asr_text>. Do not replace it with language Norwegian Bokmål<asr_text> or place style prompts in the system message. The tokenizer uses left padding, EOS token 151645, and pad token 151643. See the direct parent model card for the complete low-level inference example.

Custom tokenizer

The model inherits the parent's remapped reserved Qwen tokenizer slots without changing model dimensions. Relevant project tokens include <verbatim> (151677), <reading_optimised> (151678; reserved, since empty prompt is used), <summary> (151679), <inaudible> (151680), <laughter> (151681), <applause> (151682), and <cough> (151683).

Limitations

• This checkpoint prioritises canonical hesitation retention. Its long-form regressions make it unsuitable as a drop-in replacement when five-minute reliability is the primary requirement.
• Training data is private and may overlap public or internal evaluation sources at the recording level; no comprehensive contamination claim is made.
• The model is intended for Norwegian Bokmål. Robustness, fairness, multilingual capability, and production readiness have not been established.
• Generic ASR pipelines that ignore the bundled template and prompt contract may produce different results.
• Generated transcripts may contain omissions, substitutions, hallucinations, offensive text, and incorrectly rendered names or numbers. Human review is required for consequential use.

Published files

This repository contains inference weights and exact configuration/processor files only. Optimizer, scheduler, RNG, Trainer state, training arguments, audio, transcripts, manifests, logs, and credentials are excluded. A compact release_provenance.json records file hashes and source-run details.