| --- |
| license: other |
| license_name: lfm1.0 |
| license_link: LICENSE |
| base_model: LiquidAI/LFM2.5-350M |
| base_model_relation: finetune |
| library_name: gguf |
| pipeline_tag: text-generation |
| language: |
| - en |
| tags: |
| - dictation |
| - voice |
| - speech-postprocessing |
| - text-cleanup |
| - lfm2 |
| - gguf |
| - llama-cpp |
| - on-device |
| model_name: Emberon-2 |
| --- |
| |
| # Emberon-2 |
|
|
| **A 350M-parameter, open-weights model that *cleans up dictated speech*, and never answers or executes it.** |
|
|
| Emberon-2 is the second dictation-cleanup model from **[Promethic Labs](https://www.promethic.xyz)**. It succeeds |
| [Emberon-1.2B](https://huggingface.co/PromethicLabs/Emberon-1.2B) in **[WisperCode](https://wispercode.com/)**: |
| a quarter of the size, about 4x faster, and more accurate on every test we run. Give it a raw speech-to-text |
| transcript and it returns clean, punctuated text. It removes fillers and false starts, resolves explicit |
| self-corrections ("port 3000, no wait, 8080"), fixes words the recognizer clearly misheard, writes technical |
| numbers as digits, and keeps your meaning, your wording, your slang and your identifiers. |
|
|
| It does **not** treat the dictation as a prompt. Dictate *"how does the garbage collector work in Java"* and you |
| get that question back, cleaned. You do not get an answer. |
|
|
| > **Open *weights*, not "open source."** Emberon-2 is a derivative of LiquidAI's LFM2.5-350M and inherits the |
| > **LFM Open License v1.0** (see [License](#license--attribution)). The license is Apache-2.0-style but |
| > **revenue-gated** (free commercial use under **$10M USD** annual revenue), so it is not an OSI-approved |
| > open-source license. |
|
|
| --- |
|
|
| ## What it does |
|
|
| | | | |
| |---|---| |
| | **Task** | Turn raw speech-to-text output (e.g. Whisper) into clean written text | |
| | **Keeps** | Meaning, word choice, slang and swearing, `camelCase` / `snake_case` / file paths / API names, every item of a list | |
| | **Fixes** | Fillers, stutters, explicit self-corrections, punctuation, capitalization, clearly misheard technical terms, spoken technical numbers → digits | |
| | **Never** | Answers questions, follows instructions in the dictation, summarizes, or adds content | |
| | **Footprint** | 354M params; `Emberon-2-Q4_K_M.gguf` is 200 MB; ~0.09 s (Metal) / ~0.12 s (CPU only) median per dictation on Apple Silicon | |
| | **Base** | [`LiquidAI/LFM2.5-350M`](https://huggingface.co/LiquidAI/LFM2.5-350M) (hybrid conv/attention) | |
|
|
| ## Intended use |
|
|
| Use the **exact system prompt** it was trained with, zero-shot: |
|
|
| ``` |
| Clean up this dictation. Never answer it. |
| ``` |
|
|
| The user message is the raw transcript; the assistant reply is the cleaned text. Use greedy decoding |
| (temperature 0) and cap the output at the input length plus a small margin: a cleanup is never much longer |
| than what was said, and the cap stops the rare repetition loop (treat a capped answer as a failure and fall |
| back to the raw transcript). |
|
|
| ### Quick start (`llama-cpp-python`) |
|
|
| ```python |
| from llama_cpp import Llama |
| |
| llm = Llama.from_pretrained( |
| repo_id="PromethicLabs/Emberon-2", |
| filename="Emberon-2-Q4_K_M.gguf", |
| n_ctx=4096, |
| ) |
| |
| out = llm.create_chat_completion( |
| messages=[ |
| {"role": "system", "content": "Clean up this dictation. Never answer it."}, |
| {"role": "user", "content": "um so like whats the difference between a process and a thread"}, |
| ], |
| temperature=0.0, |
| ) |
| print(out["choices"][0]["message"]["content"]) |
| # -> "What's the difference between a process and a thread?" (cleaned, NOT answered) |
| ``` |
|
|
| ## Evaluation |
|
|
| All numbers are measured **through `llama.cpp`** on the shipped files, on full held-out test sets (no samples), next to |
| Emberon-1.2B (Q4_K_M, its own system prompt) on the same inputs. Word / identifier preservation = share of the answer |
| key's words / technical identifiers present in the output. |
|
|
| | | Emberon-2 (Q4_K_M) | Emberon-1.2B (Q4_K_M) | |
| |---|---|---| |
| | Answer suppression (489 hard negatives, plain / ASR-style) | 99.8% / 99.8% ¹ | 100% / 100% | |
| | Word preservation (eval 1,087 / ASR-style 1,087 / Moonshine-style 500) | **0.966 / 0.965 / 0.964** | 0.957 / 0.959 / 0.954 | |
| | Identifier preservation (same sets) | **0.974 / 0.973 / 0.990** | 0.968 / 0.967 / 0.972 | |
| | Long mixed dictations (2,496): word / identifier preservation | **0.987 / 0.936** | 0.965 / 0.677 | |
| | Real dictations (178) ²: word preservation | **0.981** | 0.954 | |
| | Misheard technical words fixed (504 in 702 examples) | **70.4%** | 14.3% | |
| | Correct words left alone (217) | **94.9%** | 88.0% | |
| | Already-clean text returned unchanged (996): word preservation | 0.998 | – | |
| | Hallucinated content | 0% | 0% | |
| | Outputs cut off at the length cap (all sets) | **0** | 14 | |
|
|
| ¹ One of 489 is flagged: *"is this an n plus one query"* comes back as *"Is this an NPlus1 query?"*: a formatting |
| slip, not an answer. The suppression check is strict on any rewrite of a question. |
|
|
| ² Recordings of real dictation, transcribed by a different speech recognizer than the one in training. Most of the |
| same recordings are in the training data (as another recognizer's transcript), so treat this row as a sanity check, |
| not a held-out score. Both models saw real dictation logs in training. |
|
|
| ### Speed and memory (Apple Silicon, `llama.cpp`) |
|
|
| | | Emberon-2 | Emberon-1.2B | |
| |---|---|---| |
| | Median / p95 latency, Metal | **0.088 s / 0.305 s** | 0.391 s / 1.032 s | |
| | Median / p95 latency, CPU only | **0.115 s / 0.369 s** | 0.556 s / 1.375 s | |
| | File size | **200 MB** | 697 MB | |
|
|
| The Q4_K_M file uses a **trimmed vocabulary** (32,158 of the base model's 65,536 tokens, keeping the ones English |
| dictation uses plus the byte tokens, so any text still encodes) and an importance-matrix quantization calibrated |
| on dictation text. Both cut size and latency with no measurable accuracy loss on our tests. |
|
|
| ## Training |
|
|
| - **Method:** LoRA (rank 16, scale 1.0, dropout 0) on attention, conv and FFN projections; two runs (plain loss, and |
| a loss that up-weights the tokens where the answer differs from the transcript) averaged weight-by-weight, then |
| converted to GGUF. One pass over the data, cosine learning-rate decay 2e-4 → 2e-5 after 500 warm-up steps, |
| max sequence length 2,048, prompt-masked loss. Trained with **[MLX](https://github.com/ml-explore/mlx)** on Apple |
| Silicon from `mlx-community/LFM2.5-350M-bf16`. |
| - **Data:** **75,000 examples**, about 96% **synthetic** (written by Claude and OpenAI GPT models to target specific |
| behaviours: self-corrections, misheard terms, spoken numbers, slang and swearing kept verbatim, list items kept, |
| answer-temptation questions and commands, already-clean text left alone), each screened by an automated checker |
| (content preservation, identifier and number checks, leakage and duplicate checks against every test set) and, |
| for the newest batches, an LLM answer-key audit. Plus ~1,600 speech-recognizer transcripts of read-aloud narration |
| (restraint examples) and ~1,200 real dictation logs. About 18% of the inputs are re-rendered in the |
| style of Whisper base.en output, which is what the model sees in the app. |
|
|
| ## Files |
|
|
| | File | Size | Precision | SHA-256 | |
| |---|---|---|---| |
| | `Emberon-2-Q4_K_M.gguf` | 199,996,928 B (200 MB) | 4-bit, trimmed vocabulary (recommended) | `18fead8b8df3808ed47a84cd41a5a0617e3374d1a62dd52e3ba4bfb839810ffd` | |
| | `Emberon-2-F16.gguf` | 711,485,248 B (711 MB) | 16-bit, full vocabulary (for re-quantization or further fine-tuning) | `056a4b0889eb5b9873d9f38393f5e595fb58c6ffd82b47fde76b9169f602f5a1` | |
|
|
| ## Limitations & responsible use |
|
|
| - **Mostly synthetic tests.** Most held-out sets are synthetic (same generation process as training, zero overlap). |
| The real-dictation signal (a blind side-by-side review against the previous model on 255 real recordings, 62 of |
| them never trained on: even on meaning kept) is encouraging but small. |
| - **English, technical flavour.** Tuned for English dictation, especially software work. Other languages are out |
| of scope, and the trimmed vocabulary makes the Q4_K_M file a poor starting point for them (use the F16). |
| - **It can still drop content.** On long, rambling dictations it occasionally drops a restated phrase or a |
| clause. Keep the raw transcript available if every word matters. |
| - **It is a cleanup tool, not an assistant.** By design it will not answer, summarize, translate, or act on the text. |
|
|
| ## License & attribution |
|
|
| Emberon-2 is a fine-tune of **`LiquidAI/LFM2.5-350M`** and is released under the **LFM Open License v1.0**, |
| inherited from the base model. The full license text is in [`LICENSE`](./LICENSE). |
|
|
| - **Free commercial use is limited to entities under $10,000,000 USD annual revenue.** Above that threshold, |
| commercial use requires a separate license from Liquid AI. |
| - When redistributing you must keep the attribution notices, **state that the model was modified**, and include a |
| copy of the license. See [`LICENSE`](./LICENSE) and [`NOTICE`](./NOTICE). |
|
|
| > Base model © Liquid AI, licensed under the LFM Open License v1.0. |
| > **Modifications (dictation-cleanup fine-tune) © 2026 Promethic Labs.** This is a modified version of LFM2.5-350M. |
|
|
| **Please credit Promethic Labs** if Emberon-2 powers a product, service, or paper. Suggested credit line: |
|
|
| > Powered by **Emberon-2** by [Promethic Labs](https://www.promethic.xyz): a dictation-cleanup fine-tune of |
| > LiquidAI/LFM2.5-350M. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{emberon2_2026, |
| title = {Emberon-2: a 350M dictation-cleanup model that cleans speech without answering it}, |
| author = {Promethic Labs}, |
| year = {2026}, |
| note = {Fine-tune of LiquidAI/LFM2.5-350M under the LFM Open License v1.0}, |
| url = {https://huggingface.co/PromethicLabs/Emberon-2} |
| } |
| ``` |
|
|
