134 lines
4.0 KiB
Markdown
134 lines
4.0 KiB
Markdown
---
|
|
base_model: unsloth/qwen3-4b-unsloth-bnb-4bit
|
|
tags:
|
|
- text-generation-inference
|
|
- transformers
|
|
- unsloth
|
|
- qwen3
|
|
license: apache-2.0
|
|
language:
|
|
- en
|
|
---
|
|
|
|
# OpenFable-4B
|
|
|
|
> *"The US banned Fable 5 outside America. I'm outside America. So I made my own."*
|
|
|
|
**OpenFable-4B** is a fine-tune of [Qwen3-4B](https://huggingface.co/Qwen/Qwen3-4B) designed to replicate the conversational style, reasoning depth, and structured output quality of Claude Fable 5 — built entirely from scratch by a solo developer in Brazil.
|
|
|
|
This is not a generic instruction-tuned model. It's a deliberate attempt to bring Fable-style responses to the open-source ecosystem, with a custom-built dataset and a personality baked into the chat template.
|
|
|
|
---
|
|
|
|
## What makes it different
|
|
|
|
- **Style-first fine-tune** — trained to match Claude Fable 5's tone: direct, warm, structured, and non-verbose
|
|
- **Custom dataset** — ~300 hand-curated examples across coding, math, agentic planning, and cybersecurity. No public synthetic datasets that leak CoT preambles
|
|
- **Custom chat template** — default system prompt embedded in `tokenizer_config.json`: *"You are OpenFable, created by SynastrIA Networks"*
|
|
- **GGUF quantized** — Q4_K_M, ready for local inference via llama.cpp, LM Studio, PocketPal, or Jan
|
|
|
|
---
|
|
|
|
## Benchmarks
|
|
|
|
### MMLU — Zero-shot (no few-shot)
|
|
|
|
OpenFable-4B was evaluated on MMLU with zero-shot prompting, achieving an overall score of **68.48%**.
|
|
|
|
|
|
|
|
Strongest in Social Sciences. Weakest in Humanities — expected given the dataset skew toward technical and reasoning tasks.
|
|
|
|
---
|
|
|
|
### GSM8K — Comparison with 4B-class models
|
|
|
|
OpenFable-4B holds its own against the competitive 4B landscape on grade-school math reasoning:
|
|
|
|
|
|
|
|
OpenFable matches top-tier 4B models on math reasoning despite being a style fine-tune, not a reasoning-optimized model. The base Qwen3-4B it's built on scores ~76% — OpenFable closes that gap significantly through LoRA training.
|
|
|
|
---
|
|
|
|
## Model details
|
|
|
|
| Property | Value |
|
|
|---|---|
|
|
| Base model | Qwen/Qwen3-4B |
|
|
| Fine-tuning method | LoRA (via Unsloth) |
|
|
| Dataset size | ~300 examples |
|
|
| Quantization | Q4_K_M (GGUF) |
|
|
| Context length | 32768 |
|
|
| Language | English |
|
|
| License | Apache 2.0 |
|
|
|
|
---
|
|
|
|
## Usage
|
|
|
|
### llama.cpp
|
|
|
|
```bash
|
|
./llama-cli \
|
|
-m OpenFable-4B-Q4_K_M.gguf \
|
|
-p "You are OpenFable, created by SynastrIA Networks." \
|
|
--ctx-size 4096 \
|
|
-i
|
|
```
|
|
|
|
### Python (llama-cpp-python)
|
|
|
|
```python
|
|
from llama_cpp import Llama
|
|
|
|
llm = Llama(
|
|
model_path="OpenFable-4B-Q4_K_M.gguf",
|
|
n_ctx=4096,
|
|
chat_format="chatml",
|
|
)
|
|
|
|
response = llm.create_chat_completion(
|
|
messages=[
|
|
{"role": "system", "content": "You are OpenFable, created by SynastrIA Networks."},
|
|
{"role": "user", "content": "Explain how LoRA fine-tuning works."},
|
|
]
|
|
)
|
|
|
|
print(response["choices"][0]["message"]["content"])
|
|
```
|
|
|
|
### LM Studio / Jan / PocketPal
|
|
|
|
Download the `.gguf` file and load it directly. The system prompt is already embedded in the tokenizer config — no manual setup required.
|
|
|
|
---
|
|
|
|
## Downloads
|
|
|
|
| Format | Link |
|
|
|---|---|
|
|
| GGUF (Q4_K_M) | [gustajunq/OpenFable-4B-GGUF](https://huggingface.co/gustajunq/OpenFable-4B-GGUF) |
|
|
| Org page | [SynastrIA Networks on HuggingFace](https://huggingface.co/SynastrIA-Networks) |
|
|
|
|
---
|
|
|
|
## Known limitations
|
|
|
|
- Humanities performance lags behind other categories (~59.5% MMLU) — reflective of dataset composition
|
|
- Style fine-tune, not RLHF-aligned — may occasionally drift on edge-case prompts
|
|
- Not optimized for multilingual use — English only
|
|
|
|
---
|
|
|
|
## About
|
|
|
|
Built by [Gustavo](https://huggingface.co/gustajunq) at [SynastrIA Networks](https://huggingface.co/SynastrIA-Networks) — a one-person AI startup from Brazil.
|
|
|
|
OpenFable is part of the broader SynastrIA ecosystem, which includes [Lucian](https://github.com/synastriadev), an AI agent platform for creators.
|
|
|
|
Follow the build-in-public journey: [@synastriadev](https://tiktok.com/@synastriadev) · [@openfable](https://tiktok.com/@openfable)
|
|
|
|
---
|
|
|
|
*V2 — June 2026* |