harrier77/LFM2.5_350M-ITA

LFM2.5-350M-ITA

Supervised Fine-Tuned (SFT) version of LiquidAI/LFM2.5-350M, optimized for Italian language tasks with a focus on text summarization.

Model Details

Attribute	Value
Base Model	LiquidAI/LFM2.5-350M
Architecture	Hybrid: 10 double-gated LIV convolution blocks + 6 GQA blocks
Parameters	350M
Context Length	32,768 tokens
Vocabulary Size	65,536
Training Budget	28T tokens (pre-training)
Languages	English, Arabic, Chinese, French, German, Japanese, Korean, Portuguese, Spanish (+ Italian fine-tuning)

Fine-Tuning Overview

This model was fine-tuned using Supervised Fine-Tuning (SFT) on a mixture of:

1. Italian Summarization Data - Custom Italian dataset for text summarization tasks
2. Alpaca-style Instruction Data - General instruction-following examples

The fine-tuning was performed to enhance the model's ability to:

• Generate high-quality Italian text summaries
• Follow instructions in Italian
• Better serve Italian-speaking users

Training Configuration

• Framework: TRL (Transformer Reinforcement Learning)
• Method: LoRA (Low-Rank Adaptation) / Full fine-tuning
• Training Steps: [Add your training steps]
• Batch Size: [Add your batch size]
• Learning Rate: [Add your learning rate]

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "harrier77/LFM2.5_350M-ITA"
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", dtype="bfloat16")
tokenizer = AutoTokenizer.from_pretrained(model_id)

# For chat-based interaction
messages = [
    {"role": "user", "content": "Riassumi il seguente testo: " + your_italian_text}
]

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

output = model.generate(
    input_ids,
    do_sample=True,
    temperature=0.1,
    top_k=50,
    repetition_penalty=1.05,
    max_new_tokens=512
)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Limitations

• This is a 350M parameter model - smaller than most popular LLMs
• Fine-tuned primarily on Italian summarization data; performance may vary for other tasks
• Knowledge cutoff: mid-2024 (inherited from base model)
• May not always produce accurate or factually correct summaries

License

Inherited from base model: lfm1.0

Citation

@article{lfm2.5-350m-ita,
  author = {harrier77},
  title = {LFM2.5-350M-ITA: Italian Fine-Tuned LFM2.5-350M},
  year = {2025},
  note = {Supervised Fine-Tuning on Italian summarization + Alpaca data}
}

---

This model card was automatically generated based on the base model LFM2.5-350M by Liquid AI.