70 lines
1.8 KiB
Markdown
70 lines
1.8 KiB
Markdown
---
|
||
language:
|
||
- en
|
||
license: apache-2.0
|
||
tags:
|
||
- causal-lm
|
||
- llama
|
||
- sable
|
||
- small-language-model
|
||
pipeline_tag: text-generation
|
||
---
|
||
|
||
# 🖤 Sable-Mini-30M (Preview)
|
||
|
||
**Sable-Mini** is a 30M parameter causal language model trained on approximately **3–4B tokens**. It is the flagship of the Sable tiny model family, balancing compact size with capable text generation.
|
||
|
||
> ⚠️ **This is a preview release** — expect improvements in future versions.
|
||
|
||
## 📊 Benchmark Results
|
||
|
||
Evaluated with the [LM Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness):
|
||
|
||
| Benchmark | Accuracy | Metric |
|
||
| :--- | :---: | :---: |
|
||
| **BoolQ** | **57.5%** | `acc` |
|
||
| **PIQA** | **57.8%** | `acc_norm` |
|
||
| **WinoGrande** | **53.3%** | `acc` |
|
||
| **ARC-Easy** | **39.0%** | `acc_norm` |
|
||
| **HellaSwag** | **28.1%** | `acc_norm` |
|
||
| **Lambada** | **11.7%** | `acc` |
|
||
|
||
|
||
|
||
## 🚀 Quick Start
|
||
|
||
```python
|
||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
||
tokenizer = AutoTokenizer.from_pretrained("yourorg/sable-mini-30m-preview")
|
||
model = AutoModelForCausalLM.from_pretrained("yourorg/sable-mini-30m-preview")
|
||
|
||
inputs = tokenizer("The future of tiny models is", return_tensors="pt")
|
||
outputs = model.generate(**inputs, max_new_tokens=50)
|
||
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
||
```
|
||
|
||
## 🏗️ Model Details
|
||
|
||
| Property | Value |
|
||
|----------|-------|
|
||
| Parameters | 30M |
|
||
| Architecture | Llama-style decoder |
|
||
| Training tokens | ~3–4B |
|
||
| Context length | ← add yours |
|
||
| Vocab size | ← add yours |
|
||
|
||
## ⚠️ Limitations
|
||
|
||
- Preview release — not recommended for production use
|
||
- Limited knowledge due to small parameter count
|
||
- May produce biased or incorrect outputs
|
||
|
||
## 📜 License
|
||
|
||
Apache 2.0 — free for commercial and research use.
|
||
|
||
---
|
||
|
||
*Part of the Sable family 🖤 — tiny models, sharp minds.*
|