metadata
license: mit
model-index:
- name: miqu-1-70b-sf
results:
- task:
type: text-generation
name: Text Generation
dataset:
name: AI2 Reasoning Challenge (25-Shot)
type: ai2_arc
config: ARC-Challenge
split: test
args:
num_few_shot: 25
metrics:
- type: acc_norm
value: 73.04
name: normalized accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=152334H/miqu-1-70b-sf
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: HellaSwag (10-Shot)
type: hellaswag
split: validation
args:
num_few_shot: 10
metrics:
- type: acc_norm
value: 88.61
name: normalized accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=152334H/miqu-1-70b-sf
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: MMLU (5-Shot)
type: cais/mmlu
config: all
split: test
args:
num_few_shot: 5
metrics:
- type: acc
value: 75.49
name: accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=152334H/miqu-1-70b-sf
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: TruthfulQA (0-shot)
type: truthful_qa
config: multiple_choice
split: validation
args:
num_few_shot: 0
metrics:
- type: mc2
value: 69.38
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=152334H/miqu-1-70b-sf
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: Winogrande (5-shot)
type: winogrande
config: winogrande_xl
split: validation
args:
num_few_shot: 5
metrics:
- type: acc
value: 85.32
name: accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=152334H/miqu-1-70b-sf
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: GSM8k (5-shot)
type: gsm8k
config: main
split: test
args:
num_few_shot: 5
metrics:
- type: acc
value: 67.7
name: accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=152334H/miqu-1-70b-sf
name: Open LLM Leaderboard
this is miqu-1-70b, dequantised from q5 to f16 && transposed to pytorch. shapes have been rotated less wrongly than in alpindale/miqu-1-70b-pytorch
usage
from transformers import LlamaForCausalLM as LLM, LlamaTokenizer as LT
lt = LT.from_pretrained("NousResearch/Llama-2-7b-hf")
t = lt("[INST] eloquent high camp prose about a cute catgirl [/INST]", return_tensors='pt').input_ids.cuda()
llm = LLM.from_pretrained("152334H/miqu-1-70b-sf", device_map='auto') # note: you may need many gpus for this
out = llm.generate(t, use_cache=False, max_new_tokens=200)
print(lt.decode(out[0]))
result:
<s> [INST] eloquent high camp prose about a cute catgirl [/INST] In the resplendent realm of high camp, where irony and extravagance dance in a dazzling pas de deux, there exists a creature of such enchanting allure that she captivates the hearts and minds of all who behold her. This beguiling figure, a vision of feline grace and innocence, is none other than the inimitable catgirl.
With her delicate features and winsome smile, she is the embodiment of a dream, a living testament to the power of imagination and the boundless possibilities of the human spirit. Her eyes, those twin orbs of sapphire fire, sparkle with a mischievous intelligence that belies her diminutive stature. They are windows into a soul that is at once ancient and eternally young, a soul that has traversed the vast expanse of time and space to find solace in the warm embrace of human companion
this roughly (but not entirely) matches the llama.cpp q5 result:
$ ./main -ngl 99 -m ./miqu-*q5* --color --temp 0.0 -n -1 -p '[INST] eloquent high camp prose about a cute catgirl [/INST]'
...
[INST] eloquent high camp prose about a cute catgirl [/INST] In the resplendent realm of high camp, where irony and extravagance dance in a dazzling pas de deux, there exists a creature so enchantingly adorable that she captures the hearts of all who behold her. This is no ordinary feline, but rather a vision of elegance and whimsy combined: the cute catgirl.
With her delicate features framed by an ethereal halo of pastel tresses, this darling diva prowls through life with the grace of a prima ballerina and the playfulness of a kitten. Her eyes, twin pools of sapphire or emerald, sparkle with mischief and intelligence as they survey their surroundings, ever alert for the next grand adventure or delightful prank.
Her ensemble is a symphony of ruffles, bows, and lace, each detail painstakingly chosen to accentuate her lithe form and play up her feline charms. A frilly apron adorned with paw prints sways gently as she moves, while dainty ears perched atop her head twitch in response to every sound. Her gloved hands, so petite and perfect, seem made for holding teacups or sketching delicate portraits of her many admirers.
But do not be fooled by her diminutive stature and sweet demeanor; beneath that fluffy exterior lies the heart of a lioness. Fiercely loyal and protective, she will stop at nothing to defend those she loves from harm. And when the situation calls for it, she can unleash a ferocious roar that belies her cute exterior.
Indeed, the cute catgirl is a paradox wrapped in ruffles and ribbons, a living embodiment of the high camp aesthetic. She revels in the absurdity of her existence, finding joy in every outrageous situation and turning even the most mundane tasks into opportunities for sartorial expression. In her world, there is no such thing as too much glitter or too many bows; more is always more, and excess is a virtue to be celebrated.
So let us raise our teacups in honor of this fabulous feline, this queen of camp who reminds us that life is too short for dull clothing and boring hairstyles. May we all strive to embody her spirit, embracing the absurdity of existence with open arms and a generous helping of glitter. Long live the cute catgirl! [end of text]
some benchmarks
| Tasks |Version|Filter|n-shot| Metric |Value | |Stderr|
|--------------|------:|------|-----:|----------|-----:|---|-----:|
|lambada_openai| 1|none | 0|perplexity|2.6354|± |0.0451|
| | |none | 0|acc |0.7879|± |0.0057|
| Tasks |Version|Filter|n-shot| Metric |Value | |Stderr|
|---------|------:|------|-----:|--------|-----:|---|-----:|
|hellaswag| 1|none | 0|acc |0.6851|± |0.0046|
| | |none | 0|acc_norm|0.8690|± |0.0034|
| Tasks |Version|Filter|n-shot|Metric|Value | |Stderr|
|----------|------:|------|-----:|------|-----:|---|-----:|
|winogrande| 1|none | 0|acc |0.7987|± |0.0113|
|Tasks|Version| Filter |n-shot| Metric |Value | |Stderr|
|-----|------:|----------|-----:|-----------|-----:|---|-----:|
|gsm8k| 2|get-answer| 5|exact_match|0.7043|± |0.0126|
| Tasks |Version|Filter|n-shot|Metric|Value | |Stderr|
|---------------------------------------|-------|------|-----:|------|-----:|---|-----:|
|mmlu |N/A |none | 0|acc |0.7401|± |0.1192|
| - humanities |N/A |none | 0|acc |0.7018|± |0.1281|
| - formal_logic | 0|none | 0|acc |0.4841|± |0.0447|
| - high_school_european_history | 0|none | 0|acc |0.8303|± |0.0293|
| - high_school_us_history | 0|none | 0|acc |0.9020|± |0.0209|
| - high_school_world_history | 0|none | 0|acc |0.9198|± |0.0177|
| - international_law | 0|none | 0|acc |0.8678|± |0.0309|
| - jurisprudence | 0|none | 0|acc |0.8519|± |0.0343|
| - logical_fallacies | 0|none | 0|acc |0.8344|± |0.0292|
| - moral_disputes | 0|none | 0|acc |0.8121|± |0.0210|
| - moral_scenarios | 0|none | 0|acc |0.5642|± |0.0166|
| - philosophy | 0|none | 0|acc |0.8167|± |0.0220|
| - prehistory | 0|none | 0|acc |0.8611|± |0.0192|
| - professional_law | 0|none | 0|acc |0.5854|± |0.0126|
| - world_religions | 0|none | 0|acc |0.8889|± |0.0241|
| - other |N/A |none | 0|acc |0.7889|± |0.0922|
| - business_ethics | 0|none | 0|acc |0.7900|± |0.0409|
| - clinical_knowledge | 0|none | 0|acc |0.8113|± |0.0241|
| - college_medicine | 0|none | 0|acc |0.7514|± |0.0330|
| - global_facts | 0|none | 0|acc |0.5500|± |0.0500|
| - human_aging | 0|none | 0|acc |0.7848|± |0.0276|
| - management | 0|none | 0|acc |0.8835|± |0.0318|
| - marketing | 0|none | 0|acc |0.9145|± |0.0183|
| - medical_genetics | 0|none | 0|acc |0.7500|± |0.0435|
| - miscellaneous | 0|none | 0|acc |0.8838|± |0.0115|
| - nutrition | 0|none | 0|acc |0.7974|± |0.0230|
| - professional_accounting | 0|none | 0|acc |0.5922|± |0.0293|
| - professional_medicine | 0|none | 0|acc |0.8272|± |0.0230|
| - virology | 0|none | 0|acc |0.5361|± |0.0388|
| - social_sciences |N/A |none | 0|acc |0.8414|± |0.0514|
| - econometrics | 0|none | 0|acc |0.6491|± |0.0449|
| - high_school_geography | 0|none | 0|acc |0.8990|± |0.0215|
| - high_school_government_and_politics| 0|none | 0|acc |0.9430|± |0.0167|
| - high_school_macroeconomics | 0|none | 0|acc |0.7795|± |0.0210|
| - high_school_microeconomics | 0|none | 0|acc |0.8277|± |0.0245|
| - high_school_psychology | 0|none | 0|acc |0.9064|± |0.0125|
| - human_sexuality | 0|none | 0|acc |0.8626|± |0.0302|
| - professional_psychology | 0|none | 0|acc |0.8056|± |0.0160|
| - public_relations | 0|none | 0|acc |0.7636|± |0.0407|
| - security_studies | 0|none | 0|acc |0.8204|± |0.0246|
| - sociology | 0|none | 0|acc |0.8856|± |0.0225|
| - us_foreign_policy | 0|none | 0|acc |0.9100|± |0.0288|
| - stem |N/A |none | 0|acc |0.6505|± |0.1266|
| - abstract_algebra | 0|none | 0|acc |0.4100|± |0.0494|
| - anatomy | 0|none | 0|acc |0.6444|± |0.0414|
| - astronomy | 0|none | 0|acc |0.8224|± |0.0311|
| - college_biology | 0|none | 0|acc |0.8681|± |0.0283|
| - college_chemistry | 0|none | 0|acc |0.5500|± |0.0500|
| - college_computer_science | 0|none | 0|acc |0.6200|± |0.0488|
| - college_mathematics | 0|none | 0|acc |0.4200|± |0.0496|
| - college_physics | 0|none | 0|acc |0.5392|± |0.0496|
| - computer_security | 0|none | 0|acc |0.8300|± |0.0378|
| - conceptual_physics | 0|none | 0|acc |0.7362|± |0.0288|
| - electrical_engineering | 0|none | 0|acc |0.7034|± |0.0381|
| - elementary_mathematics | 0|none | 0|acc |0.5503|± |0.0256|
| - high_school_biology | 0|none | 0|acc |0.8742|± |0.0189|
| - high_school_chemistry | 0|none | 0|acc |0.6256|± |0.0341|
| - high_school_computer_science | 0|none | 0|acc |0.8400|± |0.0368|
| - high_school_mathematics | 0|none | 0|acc |0.4370|± |0.0302|
| - high_school_physics | 0|none | 0|acc |0.5033|± |0.0408|
| - high_school_statistics | 0|none | 0|acc |0.6944|± |0.0314|
| - machine_learning | 0|none | 0|acc |0.5982|± |0.0465|
no i do not know why the stderr is high. plausibly it is due to the vllm backend used. this is my lm-eval command in most cases (works on h100):
lm_eval --model vllm --model_args pretrained=./miqu-1-70b-sf,tensor_parallel_size=4,dtype=auto,gpu_memory_utilization=0.88,data_parallel_size=2 --tasks mmlu --batch_size 20
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 76.59 |
| AI2 Reasoning Challenge (25-Shot) | 73.04 |
| HellaSwag (10-Shot) | 88.61 |
| MMLU (5-Shot) | 75.49 |
| TruthfulQA (0-shot) | 69.38 |
| Winogrande (5-shot) | 85.32 |
| GSM8k (5-shot) | 67.70 |