README.md

---
license: apache-2.0
datasets:
- roneneldan/TinyStories
language:
- en
pipeline_tag: text-generation
library_name: transformers
tags:
- sorbet
- sorbet-mini
---

# Sorbet Mini Experimental

This model isn't meant to be "good", *yet.* The whole point of Sorbet Mini
Experimental is to have a base fast enough that you can iterate on
without burning hours per experiment. It exists
to make the bigger experiments cheaper.

## What makes this so important for Sorbet

**Sorbet Mini Experimental was trained on 150M tokens, in 12 minutes, with one RTX 5060 Ti.**

What matters is that a full pretrain run in twelve minutes means every change
 on the Sorbet line can be tested quickly without burning hours on a larger model.
 
 ## What's next for Sorbet Mini
 Since Sorbet Mini is so cheap to train, it's a no brainer to keep training it. TinyStories was used to target basic language coherence as a starting point. Eventually, the full release, Sorbet Mini, will release and hopefully perform closer to other similarly sized models.

## What it is

- **Arch:** Qwen2ForCausalLM (native in transformers and llama.cpp)
- **Shape:** h192 × 8 layers, heads 6 (dim 32), GQA kv=1, inter 576, tied embeddings
- **Vocab:** 8192 (same tokenizer as the sorbet-25m family)
- **Params:** 4,920,512 total | bf16 ≈ 9.9 MB | Q8_0 ≈ 5.2 MB
- **Context:** 256 train / up to 512 inference

## Training recipe

## Architecture graph

<a href="https://hfviewer.com/CodeSoft/sorbet-mini-experimental?utm_source=huggingface&amp;utm_medium=embedded_model_card&amp;utm_campaign=CodeSoft_sorbet-mini-experimental_card" target="_blank" rel="noopener">
  <img
    src="https://hfviewer.com/api/card.svg?source=CodeSoft%2Fsorbet-mini-experimental&amp;granularity=0"
    alt="Architecture graph for CodeSoft/sorbet-mini-experimental. Open in hfviewer"
    width="100%"
  />
</a>

| knob | value |
|---|---|
| tokens | 149,999,872 (~37× Chinchilla) |
| steps | 1144 @ 512 seqs/step (seq 256) |
| data | TinyStories |
| precision | bf16, 8-bit AdamW |
| optimizer | AdamW lr 3e-4 → 1e-5 cosine, wd 0.1 (no decay on emb/norm), grad clip 1.0 |
| hardware | RTX 5060 Ti 16GB |

Result: train loss 8.13 → 2.81, val perplexity 3595 → 19.47.

## Run it

```bash
# very close to f16 (recommended)
llama-completion -m sorbet-mini-experimental-q8_0.gguf \
  -p "Once upon a time," -n 128 --temp 0.8 --top-p 0.95

# reference full-precision build
llama-cli -m sorbet-mini-experimental-f16.gguf -p "Hello, " -n 32
```

## Run the safetensors (transformers)

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tok = AutoTokenizer.from_pretrained("CodeSoft/sorbet-mini-experimental")
model = AutoModelForCausalLM.from_pretrained("CodeSoft/sorbet-mini-experimental", dtype=torch.bfloat16)
# model.to("cuda")  # if you have a GPU

prompt = "Once upon a time,"
ids = tok(prompt, return_tensors="pt").input_ids
out = model.generate(ids, max_new_tokens=128, do_sample=True,
                     temperature=0.8, top_p=0.95)
print(tok.decode(out[0], skip_special_tokens=True))
```