Tiny-LLM 25M NQ SFT

This repository contains a 25M-parameter decoder-only language model fine-tuned with supervised fine-tuning (SFT) on the sentence-transformers/natural-questions dataset.

Model Summary

  • Base model: custom Tiny-LLM 25M decoder-only model
  • Export format: Hugging Face Transformers-compatible causal language model
  • Fine-tuning method: Unsloth LoRA / QLoRA, merged into a standalone model
  • Training dataset: sentence-transformers/natural-questions
  • Task: lightweight question-answering style text generation

Prompt Format

User: {question}
Assistant:

The model was fine-tuned to continue from that prompt into an answer.

Training Details

Dataset

  • Source: sentence-transformers/natural-questions
  • Total usable records: 100,231
  • Training examples: 98,227
  • Validation examples: 2,004
  • Split seed: 1337
  • Answer handling: verbatim answers with light normalization

Fine-Tuning Setup

  • Framework: Unsloth
  • Max sequence length: 512
  • Epochs: 2
  • Effective batch size: 32
  • Final artifact: merged Hugging Face model

Intended Use

This model is intended for:

  • lightweight QA experiments
  • Tiny-LLM fine-tuning experiments
  • small-scale text generation testing
  • research and prototyping

Limitations

  • This is a very small 25M model and may produce inaccurate, incomplete, or hallucinated answers.
  • The model was fine-tuned on QA-style prompt-completion data and may not generalize well outside that format.
  • The model is not suitable for high-stakes, safety-critical, legal, financial, or medical use.

Example Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "YOUR_USERNAME/tiny-llm-25m-nq-sft"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

prompt = "User: what is gravity?\nAssistant:"
inputs = tokenizer(prompt, return_tensors="pt")

outputs = model.generate(
    **inputs,
    max_new_tokens=120,
    temperature=0.2,
    top_p=0.9,
    do_sample=True,
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Downloads last month
14
Safetensors
Model size
25M params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Dataset used to train razor5050/tinyllm-LoRA-Finetuned-25million