UNA-ThePitbull 21.4B v2

Introducing the best LLM in the industry. Nearly as good as a 70B, just a 21.4B based on saltlux/luxia-21.4b-alignment-v1.0

This model has not been poisoned to score high and be useless. We release him becaues its the real deal of EQ & IQ all together in a crazy powerful smart and conversational model.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Llamacpp imatrix Quantizations of UNA-ThePitbull-21.4B-v2

Using llama.cpp release b3001 for quantization.

Original model: https://huggingface.co/fblgit/UNA-ThePitbull-21.4B-v2

All quants made using imatrix option with dataset from here

Prompt format

<|im_start|>system
{system_prompt}<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

Download a file (not the whole branch) from below:

Downloading using huggingface-cli

First, make sure you have hugginface-cli installed:

pip install -U "huggingface_hub[cli]"

Then, you can target the specific file you want:

huggingface-cli download bartowski/UNA-ThePitbull-21.4B-v2-GGUF --include "UNA-ThePitbull-21.4B-v2-Q4_K_M.gguf" --local-dir ./

If the model is bigger than 50GB, it will have been split into multiple files. In order to download them all to a local folder, run:

huggingface-cli download bartowski/UNA-ThePitbull-21.4B-v2-GGUF --include "UNA-ThePitbull-21.4B-v2-Q8_0.gguf/*" --local-dir UNA-ThePitbull-21.4B-v2-Q8_0

You can either specify a new local-dir (UNA-ThePitbull-21.4B-v2-Q8_0) or download them all in place (./)

Which file should I choose?

A great write up with charts showing various performances is provided by Artefact2 here

The first thing to figure out is how big a model you can run. To do this, you'll need to figure out how much RAM and/or VRAM you have.

If you want your model running as FAST as possible, you'll want to fit the whole thing on your GPU's VRAM. Aim for a quant with a file size 1-2GB smaller than your GPU's total VRAM.

If you want the absolute maximum quality, add both your system RAM and your GPU's VRAM together, then similarly grab a quant with a file size 1-2GB Smaller than that total.

Next, you'll need to decide if you want to use an 'I-quant' or a 'K-quant'.

If you don't want to think too much, grab one of the K-quants. These are in format 'QX_K_X', like Q5_K_M.

If you want to get more into the weeds, you can check out this extremely useful feature chart:

llama.cpp feature matrix

But basically, if you're aiming for below Q4, and you're running cuBLAS (Nvidia) or rocBLAS (AMD), you should look towards the I-quants. These are in format IQX_X, like IQ3_M. These are newer and offer better performance for their size.

These I-quants can also be used on CPU and Apple Metal, but will be slower than their K-quant equivalent, so speed vs performance is a tradeoff you'll have to decide.

The I-quants are not compatible with Vulcan, which is also AMD, so if you have an AMD card double check if you're using the rocBLAS build or the Vulcan build. At the time of writing this, LM Studio has a preview with ROCm support, and other inference engines have specific builds for ROCm.

Want to support my work? Visit my ko-fi page here: https://ko-fi.com/bartowski

Difference V1 vs V2

On V2 we implemented a different UNA strategy and covered partially the MLP's and Attention Layers. We also performed further SFT over V1 and further DPO over V1 and we'll release some of those soon as well.

Changes

1. SFT over V1 with Replete-AI/code_bagel_hermes-2.5 at 1.0e-4 till 5.0e-5
2. DPO with: 1.0e-4 to min_lr 5.0e-5

• mlabonne/orpo-dpo-mix-40k
• jondurbin/py-dpo-v0.1

Evaluations

Can only be compared with its non-una base model: the original luxia-21.4b and ThePitbull-v1

UNA v2 (VLLM) Evaluations:

vllm (pretrained=/data/tools/mergekit/una-thepitbull-v5,dtype=bfloat16,gpu_memory_utilization=0.8,max_model_len=2048,data_parallel_size=2,tensor_parallel_size=4), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 8
|    Tasks     |Version|     Filter     |n-shot|  Metric   |Value |   |Stderr|
|--------------|------:|----------------|-----:|-----------|-----:|---|-----:|
|gsm8k         |      3|strict-match    |     5|exact_match|0.7695|±  |0.0116|+
|              |       |flexible-extract|     5|exact_match|0.7695|±  |0.0116|+
|hellaswag     |      1|none            |    10|acc        |0.8110|±  |0.0039|
|              |       |none            |    10|acc_norm   |0.9169|±  |0.0028|+
|winogrande    |      1|none            |     5|acc        |0.8777|±  |0.0092|+
|mmlu          |N/A    |none            |     0|acc        |0.6427|±  |0.0038|-
|arc_challenge |      1|none            |    25|acc        |0.7713|±  |0.0123|
|              |       |none            |    25|acc_norm   |0.7875|±  |0.0120|+
|truthfulqa_mc2|      2|none            |     0|acc        |0.7824|±  |0.0135|-
|mathqa        |      1|none            |     0|acc        |0.4037|±  | 0.009|
|              |       |none            |     0|acc_norm   |0.4034|±  | 0.009|+
|pubmedqa      |      1|none            |     0|acc        |0.7260|±  | 0.020|+
|boolq         |      2|none            |     0|acc        |0.8602|±  |0.0061|+

UNA v1 (VLLM) Evaluations

|    Tasks     |Version|     Filter     |n-shot|  Metric   |Value |   |Stderr|
|--------------|------:|----------------|-----:|-----------|-----:|---|-----:|
|gsm8k         |      3|strict-match    |     5|exact_match|0.7566|±  |0.0118|
|              |       |flexible-extract|     5|exact_match|0.7582|±  |0.0118|
|hellaswag     |      1|none            |    10|acc        |0.8168|±  |0.0039|
|              |       |none            |    10|acc_norm   |0.9188|±  |0.0027|
|winogrande    |      1|none            |     5|acc        |0.8635|±  |0.0097|
|mmlu          |    N/A|none            |     0|acc        |0.6444|±  |0.0038|
|arc_challenge |      1|none            |    25|acc        |0.7747|±  |0.0122|
|              |       |none            |    25|acc_norm   |0.7850|±  |0.0120|
|truthfulqa_mc2|      2|none            |     0|acc        |0.7902|±  |0.0134|
|mathqa        |      1|none            |     0|acc        |0.4030|±  | 0.009|
|              |       |none            |     0|acc_norm   |0.4034|±  | 0.009|
|pubmedqa      |      1|none            |     0|acc        |0.6860|±  |0.0208|
|boolq         |      2|none            |     0|acc        |0.8401|±  |0.0064|

Original (VLLM) Evaluations

|    Tasks     |Version|     Filter     |n-shot|  Metric   |Value |   |Stderr|
|--------------|------:|----------------|-----:|-----------|-----:|---|-----:|
|gsm8k         |      3|strict-match    |     5|exact_match|0.7528|±  |0.0119|
|              |       |flexible-extract|     5|exact_match|0.7521|±  |0.0119|
|hellaswag     |      1|none            |    10|acc        |0.8117|±  |0.0039|
|              |       |none            |    10|acc_norm   |0.9167|±  |0.0028|
|winogrande    |      1|none            |     5|acc        |0.8682|±  |0.0095|
|mmlu          |    N/A|none            |     0|acc        |0.6448|±  |0.0038|
|arc_challenge |      1|none            |    25|acc        |0.7688|±  |0.0123|
|              |       |none            |    25|acc_norm   |0.7730|±  |0.0122|
|truthfulqa_mc2|      2|none            |     0|acc        |0.7895|±  |0.0133|
|mathqa        |      1|none            |     0|acc        |0.4000|±  | 0.009|
|              |       |none            |     0|acc_norm   |0.4003|±  | 0.009|
|pubmedqa      |      1|none            |     0|acc        |0.6680|±  |0.0211|
|boolq         |      2|none            |     0|acc        |0.8346|±  |0.0065|

Citations

• mlabonne
• jondurbin & Replete-AI
• bartowski
• saltlux

If you use UNA models dont forget to cite:

@misc{unathepitbull21b,
  title={ThePitbull: Uniform Neural Alignment}, 
  author={Xavier Murias},
  year={2024},
  publisher = {Juanako.AI},
  journal = {HuggingFace repository},
  howpublished = {\url{https://huggingface.co/fblgit/UNA-ThePitbull-21.4-v1}},
}