CapyLake-7B-v2-laser

This model is a finetune of cognitivecomputations/WestLake-7B-v2-Laser on argilla/distilabel-capybara-dpo-7k-binarized

Process

• Realigned the chat template to ChatML
• Completed 1 Epoch
• 5e-05 learning rate
• Training time was about 2 hours on 1 H100
• Cost was ~$8

Code Example

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "macadeliccc/CapyLake-7B-v2-laser"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

text = "Create an idea for a TV show and write a short pilot script"
inputs = tokenizer(text, return_tensors="pt")

# Adding hyperparameters to the generation call
outputs = model.generate(
    **inputs,
    max_new_tokens=4096,  # Controls the maximum length of the new tokens created
    temperature=0.7,  # Adjust for creativity (lower is less random)
    top_k=50,  # Keeps the top k tokens for sampling
    top_p=0.95,  # Uses nucleus sampling with this cumulative probability
    num_return_sequences=1,  # Number of sequences to generate
    no_repeat_ngram_size=2,  # Prevents repeating n-grams to ensure diversity
    early_stopping=True  # Stops generation when all sequences reach the EOS token
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Other Capy Models

SOLAR-10.7B-Capy-v1.0 is also on the way. There could be more depending on performance!

Evaluations

Model	AGIEval	GPT4All	TruthfulQA	Bigbench	Average
CapyLake-7B-v2-laser	44.34	77.77	68.47	47.92	59.62

AGIEval

Task	Version	Metric	Value		Stderr
agieval_aqua_rat	0	acc	28.35	±	2.83
		acc_norm	25.98	±	2.76
agieval_logiqa_en	0	acc	38.86	±	1.91
		acc_norm	39.02	±	1.91
agieval_lsat_ar	0	acc	25.22	±	2.87
		acc_norm	24.35	±	2.84
agieval_lsat_lr	0	acc	50.39	±	2.22
		acc_norm	51.57	±	2.22
agieval_lsat_rc	0	acc	65.06	±	2.91
		acc_norm	63.94	±	2.93
agieval_sat_en	0	acc	78.64	±	2.86
		acc_norm	78.64	±	2.86
agieval_sat_en_without_passage	0	acc	40.78	±	3.43
		acc_norm	40.78	±	3.43
agieval_sat_math	0	acc	33.64	±	3.19
		acc_norm	30.45	±	3.11

Average: 44.34%

GPT4All

Task	Version	Metric	Value		Stderr
arc_challenge	0	acc	66.89	±	1.38
		acc_norm	67.49	±	1.37
arc_easy	0	acc	86.70	±	0.70
		acc_norm	81.90	±	0.79
boolq	1	acc	88.10	±	0.57
hellaswag	0	acc	71.45	±	0.45
		acc_norm	87.78	±	0.33
openbookqa	0	acc	39.80	±	2.19
		acc_norm	49.80	±	2.24
piqa	0	acc	82.86	±	0.88
		acc_norm	84.87	±	0.84
winogrande	0	acc	84.45	±	1.02

Average: 77.77%

TruthfulQA

Task	Version	Metric	Value		Stderr
truthfulqa_mc	1	mc1	53.98	±	1.74
		mc2	68.47	±	1.53

Average: 68.47%

Bigbench

Task	Version	Metric	Value		Stderr
bigbench_causal_judgement	0	multiple_choice_grade	59.47	±	3.57
bigbench_date_understanding	0	multiple_choice_grade	64.50	±	2.49
bigbench_disambiguation_qa	0	multiple_choice_grade	44.96	±	3.10
bigbench_geometric_shapes	0	multiple_choice_grade	22.84	±	2.22
		exact_str_match	2.79	±	0.87
bigbench_logical_deduction_five_objects	0	multiple_choice_grade	30.80	±	2.07
bigbench_logical_deduction_seven_objects	0	multiple_choice_grade	21.57	±	1.56
bigbench_logical_deduction_three_objects	0	multiple_choice_grade	56.67	±	2.87
bigbench_movie_recommendation	0	multiple_choice_grade	51.60	±	2.24
bigbench_navigate	0	multiple_choice_grade	51.00	±	1.58
bigbench_reasoning_about_colored_objects	0	multiple_choice_grade	70.35	±	1.02
bigbench_ruin_names	0	multiple_choice_grade	51.79	±	2.36
bigbench_salient_translation_error_detection	0	multiple_choice_grade	35.97	±	1.52
bigbench_snarks	0	multiple_choice_grade	79.01	±	3.04
bigbench_sports_understanding	0	multiple_choice_grade	75.66	±	1.37
bigbench_temporal_sequences	0	multiple_choice_grade	47.90	±	1.58
bigbench_tracking_shuffled_objects_five_objects	0	multiple_choice_grade	23.84	±	1.21
bigbench_tracking_shuffled_objects_seven_objects	0	multiple_choice_grade	18.00	±	0.92
bigbench_tracking_shuffled_objects_three_objects	0	multiple_choice_grade	56.67	±	2.87

Average: 47.92%

Average score: 59.62%

Elapsed time: 01:57:56