Sakura-SOLAR-Instruct

Quantizations

I also provide zipped quantization because a lot of people find gguf single download convenient. Zipped quantization is relatively smaller in size to download. After extracted, you can use the model folder as usual.

Parameter Explanation

BPW: Bits per weight of the quantized model.
Folder Size: Size of the model folder in GB when stored on disk or after extracted from zip file.
Zipped Folder Size: Size of the zipped file after downloaded.
VRAM Usage: VRAM usage of the model measured using ExLlamav2_HF and 4096 max_seq_len with Oobabooga's Text Generation WebUI.
Loss: Accuracy loss of the quantized model compared to the original model.
Description: Recommended GPU and/or settings for the model.

On the table, I included VRAM usages of both 4096 max_seq_len and 8192 max_seq_len as a comparison. Sonya-7B default max_seq_len is 8192.

Branch	BPW	Folder Size	Zipped File Size	VRAM Usage (4096/8192)	Loss	Description
2.0bpw/2.0bpw-zip	2.0BPW	2.15GB	1.80GB	3.1GB/3.5GB	~14%	For >=4GB VRAM cards with idle VRAM atleast or below 500MB (headroom for other things)
2.4bpw/2.4bpw-zip	2.4BPW	2.29GB	1.98GB	3.3GB/3.7GB	~13%	For >=4GB VRAM cards (use 4096 max_seq_len)
3.5bpw/3.5bpw-zip	3.5BPW	3.19GB	2.91GB	4.2GB/4.6GB	~7%	For >=5GB VRAM cards (use 4096 max_seq_len)
4.5bpw (main)/4.5bpw-zip	4.5BPW	4.0GB	3.77GB	5.0GB/5.3GB	~4%	For >=6GB VRAM cards with idle VRAM atleast or below 500MB (headroom for other things)
4.75bpw/4.75bpw-zip	4.75BPW	4.2GB	4.0GB	5.2GB/5.6GB	~3%	For >=6GB VRAM cards (use 4096 max_seq_len)
7.0bpw/7.0bpw-zip	7.0BPW	6.03GB	5.74GB	7GB/7.3GB	~1%	For >=8GB VRAM cards with idle VRAM atleast or below 500MB (headroom for other things)

Unfortunately I couldn't make higher quants than 7.0bpw. No error on the console, just "Killed" after creating output directory for the model.

Calibration Dataset

Exllamav2 default dataset

Prompt template: Alpaca

Below is an instruction that describes a task. Write a response that appropriately completes the request.

### Instruction:
{prompt}

### Response:

Original Info

Top 1 Performer MT-bench 🤪

WTF is This?

Sonya-7B is, at the time of writing, the #1 performing model in MT-Bench first turn, ahead of GPT-4, and overall the #2 model in MT-Bench, to the best of my knowledge. Sonya-7B should be a good all-purpose model for all tasks including assistant, RP, etc.

Sonya-7B has a similar structure to my previous model, Silicon-Maid-7B, and uses a very similar merge. It's a merge of xDAN-AI/xDAN-L1-Chat-RL-v1, Jan-Ai's Stealth v1.2, chargoddard/piano-medley-7b, NeverSleep/Noromaid-7B-v0.2, and athirdpath/NSFW_DPO_vmgb-7b. Sauce is below. Somehow, by combining these pieces, it substantially outscores any of its parents on MT-Bench.

I picked these models because:

MT-Bench normally correlates well with real world model quality and xDAN performs well on it.
Almost all models in the mix were Alpaca prompt formatted which gives prompt consistency.
Stealth v1.2 has been a magic sprinkle that seems to increase my MT-Bench scores.
I added RP models because it boosted the Writing and Roleplay benchmarks 👀

Based on the parent models, I expect this model to be used with an 8192 context window. Please use NTK scaling alpha of 2.6 to experimentally try out 16384 context.

Let me be candid: Despite the test scores, this model is NOT is a GPT killer. I think it's a very sharp model for a 7B, it probably punches way above its weight for a 7B, but it's still a 7B model. Even for a 7B model, I think it's quirky and has some weird outputs, probably due to how Frankenstein this merge is. Keep your expectations in check 😉

MT-Bench Average Turn

model	score	size
gpt-4	8.99	-
Sonya-7B	8.52	7b
xDAN-L1-Chat-RL-v1	8.34	7b
Starling-7B	8.09	7b
Claude-2	8.06	-
Silicon-Maid	7.96	7b
Loyal-Macaroni-Maid	7.95	7b
gpt-3.5-turbo	7.94	20b?
Claude-1	7.90	-
OpenChat-3.5	7.81	-
vicuna-33b-v1.3	7.12	33b
wizardlm-30b	7.01	30b
Llama-2-70b-chat	6.86	70b

The Sauce

models:
  - model: xDAN-AI/xDAN-L1-Chat-RL-v1
    parameters:
      weight: 1
      density: 1
  - model: chargoddard/piano-medley-7b
    parameters:
      weight: 0.3
  - model: jan-hq/stealth-v1.2
    parameters:
      weight: 0.2
  - model: NeverSleep/Noromaid-7b-v0.2
    parameters:
      weight: 0.2
  - model: athirdpath/NSFW_DPO_vmgb-7b
    parameters:
      weight: 0.2
merge_method: ties
base_model: mistralai/Mistral-7B-v0.1
parameters:
  density: 0.4
  int8_mask: true
  normalize: true
dtype: bfloat16

There was no additional training, finetuning, or DPO. This is a straight merger.

Prompt Template (Alpaca)

Below is an instruction that describes a task. Write a response that appropriately completes the request.

### Instruction:
{prompt}

### Response:

I found that this model performed worse with the xDAN prompt format so, despite the heavy weight of xDAN in this merger, I recommeend against its use.

Other Benchmark Stuff

########## First turn ##########

model	turn	score	size
Sonya-7B	1	9.06875	7b
gpt-4	1	8.95625	-
xDAN-L1-Chat-RL-v1	1	8.87500	7b
xDAN-L2-Chat-RL-v2	1	8.78750	30b
claude-v1	1	8.15000	-
gpt-3.5-turbo	1	8.07500	20b
vicuna-33b-v1.3	1	7.45625	33b
wizardlm-30b	1	7.13125	30b
oasst-sft-7-llama-30b	1	7.10625	30b
Llama-2-70b-chat	1	6.98750	70b

########## Second turn ##########

model	turn	score	size
gpt-4	2	9.025000	-
xDAN-L2-Chat-RL-v2	2	8.087500	30b
Sonya-7B	2	7.962500	7b
xDAN-L1-Chat-RL-v1	2	7.825000	7b
gpt-3.5-turbo	2	7.812500	20b
claude-v1	2	7.650000	-
wizardlm-30b	2	6.887500	30b
vicuna-33b-v1.3	2	6.787500	33b
Llama-2-70b-chat	2	6.725000	70b

If you'd like to replicate the MT-Bench run, please ensure that the Alpaca prompt template is applied to the model. I did this by putting "alpaca" in the model path to trigger the AlpacaAdapter.