SmolLM2-360M-parakeet-150k-FullFT-MLX-8bit

SmolLM2 at full data scale. More data improved exact match (24.5→31%) but the architecture remains dominated by the LFM-350M-Base.

Training details

Base model	HuggingFaceTB/SmolLM2-360M-Instruct
Recipe	Full fine-tune
Training data	108k pairs from rdsm/parakeet-stt-redone
Quantization	MLX 8-bit

Benchmark (100 held-out pairs, Codex judge vs GLM-5.1-FP8 references)

Metric	Score
Semantic match	71%
Exact match	31%
Serving latency	~0.67 s (LM Studio)
GLM ceiling	95%
Echo baseline	91% (do nothing)

Task

Fine-tuned to replicate GLM-5.1-FP8's transcript cleanup in MacParakeet: raw ASR text → polished, punctuated, readable text.

Prompt

System: You are a transcription formatting assistant. Follow the user's
formatting instructions exactly and return only the final formatted transcript.

User: You are a transcription cleanup assistant.

Convert the following raw transcript into polished, readable text.

Instructions:
1. Add punctuation and capitalization.
2. Split the text into natural sentences.
3. Break the text into readable paragraphs whenever the speaker moves into a new topic.
4. Prefer short paragraphs of 1 to 3 sentences.
5. Fix obvious speech-to-text errors.
6. Remove repeated words and filler sounds when unnecessary.
7. Keep the original meaning, tone, language and wording as close as possible.
8. Do not summarize, shorten, or add content.
9. Output only the final cleaned text.

Raw transcript:
{raw_transcript}

Response format: {"cleaned_text": "..."} · Temperature: 0.2