Barbet 1B
Structured SFT
A text backbone for structured editing and evidence-grounded decisions.
One full SFT epoch on Traditional Chinese task data. Explore the exact requests, generated answers, measured gains, and observed failures before building on the checkpoint.
Text in. Structured decisions out.
Work with ASR transcripts, TTS written and spoken forms, OCR candidates, and translation revisions. Requests include explicit task contracts and visible reference evidence. Audio encoders, image encoders and speech waveform generation are not included.
Read what the model actually said.
Eight inputs fixed before inference on 19 September 2026. Four structured requests contain reference text; four plain instructions test a different use pattern. These are recorded outputs, not a live demo or a benchmark.
Complete input
Raw model output
Reference answer
The gains and the cost.
Identical fixed synthetic test: 1,020 answer-loss examples and 408 generations per model. Whole-answer exact includes every field, including synthetic confidence. Parse success is not answer correctness.
| Domain | Base NLL | SFT NLL | SFT exact | Parse rate |
|---|---|---|---|---|
| ASR | 1.6994 | 0.0029 | 27/48 | 100% |
| TTS | 1.6405 | 0.0027 | 19/32 | 100% |
| OCR | 1.6231 | 0.0026 | 24/40 | 100% |
| MT | 1.5686 | 0.0024 | 13/32 | 100% |
General-text retention
Existing 6,955-row development corpus. Bits per byte (BPB) measures text prediction loss; lower is better. These are relative loss changes, not accuracy-point changes.
| Category | Base | SFT | BPB change |
|---|---|---|---|
| zh_tw_zh | 1.1399 | 1.1756 | +3.13% |
| math | 0.6890 | 0.8007 | +16.21% |
| ja_ko | 0.9680 | 0.9833 | +1.58% |
| english_general | 0.8615 | 0.8727 | +1.30% |
| code | 0.7154 | 0.9111 | +27.36% |
| multilingual | 1.1374 | 1.1679 | +2.69% |
Run a complete structured request.
The loader preserves the original BF16 / FP32 parameter mix. The example’s assistant reference is never included in the prompt.
git clone https://github.com/voidful/barbet-structured-sft.git cd barbet-structured-sft python -m venv .venv source .venv/bin/activate python -m pip install torch==2.11.0 packaging ninja setuptools wheel python -m pip install --no-build-isolation -r requirements-inference.txt python scripts/infer.py --example ocr_correction --device cuda:0
Tested: Python 3.12 · PyTorch 2.11 · CUDA 13.0 · H200. Native Mamba packages require a compatible CUDA build environment. Short CPU inference is supported.
Source, evidence, and reproducibility.
Training covered every one of the 980,000 examples once, without truncation. Public inference and synthetic-data evaluation are documented. Exact retraining additionally needs the separately controlled base checkpoint; the historical retention corpus is not distributed.
License status
Apache-2.0 covers the project code, released SFT weights, documentation and bundled examples. Third-party components and the separately published full training dataset retain their own terms. LICENSE ↗