Hi everyone,

I need help figuring out how to fully fine-tune smaller open-source language models with a new name for them (not LoRA/adapters), and also how to create my own small models from scratch (not my preference, since it requires more resources — I just want to understand the process).

My setup:

GPU: RTX 4070 Super, 12 GB VRAM

RAM: 16 GB system memory

Single GPU only

My goals:

1. Fine-tune full models under 7B parameters (preferably 0.5B–3B for easier training on my hardware).

2. Use custom datasets (my own data) and also be able to integrate additional public datasets.

3. Optionally train a small model from scratch if possible.

4. Output a full model checkpoint with updated weights (not just LoRA weights).

5. Be able to update the model’s knowledge with new data over time.(Locally or using huggingface)

What I’m looking for:

Recommendations for base models that can be fully fine-tuned with my specs.

Unsloth notebooks, scripts, or workflows that support full fine-tuning on a single GPU.

VRAM optimization tips (batch size, sequence length, gradient checkpointing, DeepSpeed/ZeRO, etc.).

Guidance or examples for building a small model from scratch.

I’ve already checked out the official Unsloth notebooks, but they’re a bit complex to adapt to my specific setup. If anyone has simplified versions or step-by-step examples for similar hardware, that would be a huge help.

If you’ve tried something like this, please share your experience, notebooks, or guides.

Thanks!