QuartetAnemoi-70B-t0.0001-b4.6-h8-exl2

Exl2 quant parameters:
Bits per weight: 4.6
Head bits: 8
Fits in 48GB with 16K of context. Try 8 bit kvcache if you want 32K.

A sequential merge using a custom algorithm (NearSwap) of:

• 152334H/miqu-1-70b-sf
• Sao10K/WinterGoddess-1.4x-70B-L2
• Aurora-Nights-70B-v1.0
• Xwin-LM-70B-V0.1

In our testing, this model seems like a storyteller, as might be expected, but the changes from this merge are extremely soft. We were impressed that, unlike most models, at the end of a story it did not often use cliches such as "In the end", "And so", "beacon of hope", etc.

Quants

Most of the popular quant formats are available now, thanks to community efforts.

Type	Misc	Author
GGUF		alchemonaut
GGUF	iMat	Nexesenex
GGUF	iMat	mradermacher
GGUF	Full Set	mradermacher
exl2	2.5bpw	llmixer
exl2	4.0bpw	llmixer
exl2	6.0bpw	llmixer
AWQ		tachyphylaxis

NearSwap Algorithm

NearSwap retains most of the weights of the base model (Miqu), but when a weight is similar between the two, it is interpolated to the secondary model value. A parameter t specifies the sameness threshold. When the distance between two values is below t, the weight from the secondary model is used.

This version of the model uses t = 0.0001. At this t, about 0.8% of weights are fully switched to the secondary model during each pass. Model quality rapidly degrades above t = 0.0025:

• t = 0.0001 (~0.8% full swap): This model
• t = 0.0003 (~2% full swap)
• t = 0.001 (~10% full swap): BoreanGale-70B
• t = 0.0025 (~18% full swap): Generates one paragraph okay, but then reverts to garbage
• t = 0.005 (~35% full swap): Garbage; semi-related word lists
• t = 0.01 (~55% full swap): Garbage; pseudorandom tokens output

For QuartetAnemoi-70B-t0.0001, the three secondary models were each merged sequentially with t = 0.0001.

NearSwap implementation:

    t: Union[float, np.ndarray],
    v0: Union[np.ndarray, torch.Tensor],
    v1: Union[np.ndarray, torch.Tensor],
...
    lweight = numpy.absolute(v0-v1)
    lweight = t / lweight
    lweight = numpy.nan_to_num(lweight, nan=1.0, posinf=1.0, neginf=1.0)
    numpy.clip(lweight, a_min=0.0, a_max=1.0, out=lweight)
    res = lerp(lweight,v0,v1)

License and Use

Since the ultimate origin of Miqu is at this time unknown beyond speculation, this model is for noncommercial research use only.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	76.86
AI2 Reasoning Challenge (25-Shot)	73.38
HellaSwag (10-Shot)	88.9
MMLU (5-Shot)	75.42
TruthfulQA (0-shot)	69.53
Winogrande (5-shot)	85.32
GSM8k (5-shot)	68.61