meeko-1-preview

Meeko 1 Preview

<p align="center"> <img src="assets/meeko.jpeg" alt="Meeko the cat" width="520"> </p>

Meeko 1 Preview is a 1.2B-parameter transcript-cleanup model. It turns raw speech-to-text output into concise written text with no system prompt.

The model is named after Meeko, the cat pictured above.

What it does

Given a raw transcript as the user message, Meeko returns only the cleaned text. Its target editing register is medium-touch written English:

• fix punctuation, capitalization, spacing, and grammar;
• remove speech fillers and stutters;
• expand casual forms such as “gonna” to “going to”;
• split run-on sentences when needed;
• format numbers, dates, times, money, percentages, symbols, and lists;
• resolve explicit self-corrections so the final value wins;
• preserve intent, facts, hedges, contrasts, identifiers, and all non-retracted content.

It is an editor, not a general-purpose chat assistant. Do not add a system prompt for the intended cleanup task.

Quick start

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "tigerisaac/meeko-1-preview"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
)

messages = [{
    "role": "user",
    "content": "um so i was gonna send it friday no wait monday morning",
}]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    generated = model.generate(
        inputs,
        max_new_tokens=128,
        do_sample=False,
    )

new_tokens = generated[0, inputs.shape[-1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))
# I was going to send it Monday morning.

GGUF (llama.cpp)

A pre-quantized Q4_K_M build is included for local inference:

• File: meeko-1-preview-Q4_K_M.gguf
• No system prompt; pass the raw transcript as the user message.

llama-server -m meeko-1-preview-Q4_K_M.gguf \
  --host 127.0.0.1 --port 8080 \
  --ctx-size 16384 --parallel 1 --reasoning off

Training

• Base model: LiquidAI/LFM2.5-1.2B-Instruct
• Method: supervised fine-tuning with Unsloth
• Epochs: 2
• Training format: one raw transcript user message followed by one cleaned assistant response; no system message
• Corpus: 11,438 curated cleanup pairs, assembled into 10,833 training and 570 validation examples after deduplication
• Label target: medium-touch written-English cleanup

The corpus includes ordinary dictation, fillers, stutters, number and symbol formatting, listification, explicit retractions, and matched contrast/hedge examples that must not be over-edited. Constructed number and retraction rows use mechanically verified labels.

License

This fine-tune is derived from LiquidAI/LFM2.5-1.2B-Instruct and is released under the applicable upstream model terms. Review the base model repository before use or redistribution.

Acknowledgments

Built with Liquid AI's LFM2.5 base model and Unsloth. Named for Meeko, who provided supervision of the non-gradient variety.