CoolFace
Modelpublic

Buttermilk03/scriber-lfm2.5-350m-polishing-de-qad-v1

sourceHugging Faceotherupdated 25d agoView on Hugging Face
0likes68downloads
Model Card

Scriber LFM2.5 350M German Post-Processing — QAD Q4_0

Training-tool attribution: These model weights are Generated Output from Praxist by Sapient Intelligence. The PRAXIST Fair Source License applies independently of the base-model license; see the pinned PRAXIST license.
Modified derivative: These weights were modified by the open-source Scriber project from LiquidAI/LFM2.5-350M-Base at revision 9960764e30892e01f29a6dc23df2533fcd8bd5ae. Liquid AI did not produce or endorse the modified model. See MODIFICATIONS.md and the included LFM Open License v1.0.

Purpose

This is a small local German live-dictation post-processor. It cleans raw speech-to-text transcripts linguistically, typographically, and structurally while preserving content, order, numbers, names, and meaning. It is not a general-purpose assistant and must not answer questions found in a transcript.

The only model artifact is:

  • —gguf/qad_q4_0/Scriber-LFM2.5-350M-Production-QAD-Q4_0.gguf
  • —size: 218,328,640 bytes
  • —SHA-256: e1ca3391d896db64df91c5ed5a02e16f5b6bbec5de81667ec99535eb7b1c0486
  • —tensor layout: 92 Q40, 1 Q6K, 55 F32 tensors

No BF16, PTQ, Q8, or other quantized product alternative is published.

Training lineage

  • —Base model: LiquidAI/LFM2.5-350M-Base, exact revision above (354,483,968 parameters).
  • —Sole source corpus: 2,000 newly generated German letter pairs. Earlier Scriber corpora and earlier SAPI, Qwen, Soniox, or historical PRAXIST inputs were not used.
  • —The initial 1,600/200/200 split was deterministically randomized before use.
  • —Production SFT then consumed all 2,000 source pairs.
  • —Mixed recovery consumed 8,000 scheduled examples constructed only from those source pairs and their deterministic short-form derivations.
  • —QAD recovery consumed 6,000 unique examples: the 2,000 original pairs plus 4,000 deterministic short-form derivations. It used a frozen merged teacher, a trainable student, completion-only forward KL, and Q4_0 fake quantization.
  • —Serialization used llama.cpp b10158 at revision f87067841bac583bc089a225382248d857791ca8.

The exact 2,000 source/target pairs, split audit, complete accepted training recipe, and lessons from the rejected runs are published in `Buttermilk03/scriber-lfm2.5-350m-polishing-de-training-v1`.

Measured behavior

The selected QAD artifact reproduced all 600 cases in the final local runtime regression exactly:

Runtime regressionCasesExactSafety rejectionProtected mutation
Long German letters200200/20000
Short identity/noise cases400400/40000
Total600600/60000

On an NVIDIA RTX 4070 Laptop GPU with llama.cpp b10158, measured median latency was about 687 ms for the long set and 196 ms for the short set. These 600 cases belong to the frozen training/validation recipe family and therefore establish runtime fidelity, not an unbiased estimate of generalization.

The separately retrained full-capacity H200 candidate reached 599/600 because it changed one protected amount. It was rejected; this repository contains the earlier 600/600 artifact.

Prompt and decoding

Use deterministic greedy decoding with at most 384 new tokens and the exact plain-completion prompt below:

text
Aufgabe: Glätte das folgende deutsche Speech-to-Text-Transkript sprachlich, typografisch und strukturell. Bewahre Inhalt, Reihenfolge, Zahlen, Namen und Bedeutung. Füge nichts hinzu, beantworte keine Fragen und gib ausschließlich die bereinigte Fassung zurück.

Transkript:
${transcript}

Bereinigte Fassung:

The prompt SHA-256 is e0ff2d5297f3d4d5ae7b8af85ea1cf52a24704bfb2e61990eab6de52b42058d8.

Safety integration

Do not expose model output without a preservation check. Scriber returns the original transcript whenever output is empty, oversized, repetitive, leaks control markers, changes protected numeric or lexical anchors, or otherwise violates the post-processing contract.

scriber-protection-policy.json contains the exact safety metadata used by Scriber. The model itself receives the raw transcript without KEEP markers; those markers must never appear in visible output.

Limitations

  • —The model is focused on German dictated letters and related structured prose.
  • —A 350M-parameter model can still omit, repeat, or alter content on unfamiliar or adversarial inputs.
  • —Names, identifiers, dates, amounts, phone numbers, addresses, and units need host-side preservation checks.
  • —CPU and GPU behavior depends on the llama.cpp build and host hardware.

License

The included LICENSE is mechanically unchanged from the pinned base-model revision. The LFM Open License v1.0 includes a Commercial Use Limitation tied to annual revenue; review the license itself before redistribution or commercial deployment.

These weights were generated with Praxist by Sapient Intelligence. The pinned PRAXIST Fair Source License requires retaining that product-name attribution for third-party availability and defines a separate commercial license threshold. On 2026-09-01, the owner explicitly confirmed aggregate annual revenue of the licensee and all affiliates of USD 0, satisfying the free-license revenue condition for this open-source work.

This model card is descriptive and is not legal advice.