MiniBARD / README.md
Naphula's picture
Update README.md
3f6f76e verified
|
Raw
History Blame Contribute Delete
3.68 kB
---
title: MiniBARD
emoji: 📉
colorFrom: red
colorTo: yellow
sdk: static
pinned: false
---
# MiniBARD
**Benchmark for Aesthetics, Roleplay, and Depth**
> That is a fantastic title. **B.A.R.D. (Benchmark for Aesthetics, Roleplay, & Depth)** feels prestigious and perfectly encapsulates the "vibe" of the models you are testing.
>
> * **Aesthetics:** Captures the Creative Writing and the "beauty" of the prose.
> * **Roleplay:** Directly addresses the RP-Bench and character immersion.
> * **Depth:** Covers the Reasoning and EQ—the model's ability to understand complex subtext and provide nuanced, multi-layered responses.
>
> This project establishes a high-precision, local benchmarking pipeline designed to compare the creative writing and roleplay capabilities of two 12B Large Language Models (Qliphoth v1a and v1b) without relying on external API keys. By utilizing a runpod 3090 GPU server and 500GB of local storage, the system bypasses massive, generic benchmarks like MMLU in favor of a targeted "Council of Judges" approach. This method uses five diverse, high-parameter "judge" models—including Gemma 3, Nemotron, and Cydonia—to evaluate model outputs across 80 complex prompts. To ensure scientific integrity, the pipeline implements a "blind" pairwise comparison that shuffles the order of responses to eliminate first-entry bias and utilizes "abliterated" judges to prevent moralizing or refusals from skewing the creative scores. The final result is a majority-rule verdict that provides a definitive win rate, offering a clear, compute-efficient data point on which model version produces superior prose and instruction-following.
The full BARD suite evaluates LLM reasoning, emotional intelligence, creative writing, and roleplay. It also removes all non-english prompts.
The BARD tool (designed for runpod) allows the user to specify either API keys or a locally hosted council of LLM judges which the benchmark output JSONs are then processed through.
```
=== PROMPT LOADING COMPLETE ===
Total English prompts loaded: 2863
- mt_bench: 80 prompts
- eq_bench: 1573 prompts
- cw_bench: 419 prompts
- rp_bench: 791 prompts
===============================
```
MiniBARD uses Bernoulli sampling with a fixed seed `(random.seed(420))` to generate perfectly diverse, representative slices of the full benchmark. This achieves high fidelity of the full BARD score while only requiring 10% of the compute power.
```
=== MINI-B.A.R.D. PROMPT LOADING COMPLETE ===
Total representative prompts loaded: 320
- mt_bench: 80 prompts
- eq_bench: 80 prompts
- cw_bench: 80 prompts
- rp_bench: 80 prompts
=============================================
```
BARD is a composite of the following benchmarks:
- [MT](https://github.com/lm-sys/FastChat)
- [EQ](https://github.com/EQ-bench/EQ-Bench)
- [CW](https://github.com/EQ-bench/creative-writing-bench)
- [RP](https://github.com/LeviTheWeasel/rp-benchmark)
Thanks [@sam-paech](https://huggingface.co/sam-paech) for releasing EQ-Bench suite.
## Examples
- [MT Only](https://cdn-uploads.huggingface.co/production/uploads/68e840caa318194c44ec2a04/udXeUI5b8QyYGKkSEx_es.png)
- [Dataset](https://huggingface.co/datasets/Naphula-Archives/qliphoth_12B_minibard_bench)
- [Example 3: Contains MT, EQ, and RP (no CW)](https://cdn-uploads.huggingface.co/production/uploads/68e840caa318194c44ec2a04/NtsvJTHaEdvmRuaBSXHSn.png)
## Current Judge Models
- Cydonia 4.3 24B
- Mag Mell 12B
- Nemotron 8B Ablit
- <s>Gemma 3 27B Ablit</s>
## Future Features Planned
- Slop tests
- Censorship tests
- [Q0 Benchmark](https://huggingface.co/Naphula/Q0_Bench) scores
- Possible representative slicing of MMLU, HellaSwag, etc.