Spaetzle-v12-7b

Spaetzle-v12-7b is a merge of the following models using LazyMergekit:

• flemmingmiguel/NeuDist-Ro-7B
• Blizado/discolm-mfto-7b-german-v0.1
• ResplendentAI/Flora_DPO_7B
• on the basis of mayflowergmbh/Wiedervereinigung-7b-dpo-laser

As expected, this is a little bit worse in general English tasks over cstr/spaetzle-v8-7b, but a tiny little bit better on German tasks, at least some: e.g. it reaches an EQ-Bench (de) score of 64.81, but only

Metric	Value
Avg.	69.36
AI2 Reasoning Challenge (25-Shot)	65.96
HellaSwag (10-Shot)	86.16
MMLU (5-Shot)	63.48
TruthfulQA (0-shot)	57.84
Winogrande (5-shot)	80.03
GSM8k (5-shot)	62.70

Model	AGIEval	GPT4All	TruthfulQA	Bigbench	Average
Spaetzle-v12-7b	42.64	74.3	58.44	44.44	54.95

AGIEval

Task	Version	Metric	Value		Stderr
agieval_aqua_rat	0	acc	24.02	±	2.69
		acc_norm	21.65	±	2.59
agieval_logiqa_en	0	acc	36.10	±	1.88
		acc_norm	37.63	±	1.90
agieval_lsat_ar	0	acc	24.35	±	2.84
		acc_norm	23.04	±	2.78
agieval_lsat_lr	0	acc	48.82	±	2.22
		acc_norm	47.25	±	2.21
agieval_lsat_rc	0	acc	60.59	±	2.98
		acc_norm	57.99	±	3.01
agieval_sat_en	0	acc	76.21	±	2.97
		acc_norm	74.76	±	3.03
agieval_sat_en_without_passage	0	acc	46.60	±	3.48
		acc_norm	45.63	±	3.48
agieval_sat_math	0	acc	37.27	±	3.27
		acc_norm	33.18	±	3.18

Average: 42.64%

GPT4All

Task	Version	Metric	Value		Stderr
arc_challenge	0	acc	59.13	±	1.44
		acc_norm	61.26	±	1.42
arc_easy	0	acc	83.67	±	0.76
		acc_norm	80.89	±	0.81
boolq	1	acc	87.83	±	0.57
hellaswag	0	acc	66.45	±	0.47
		acc_norm	84.63	±	0.36
openbookqa	0	acc	37.40	±	2.17
		acc_norm	45.80	±	2.23
piqa	0	acc	82.15	±	0.89
		acc_norm	83.13	±	0.87
winogrande	0	acc	76.56	±	1.19

Average: 74.3%

TruthfulQA

Task	Version	Metric	Value		Stderr
truthfulqa_mc	1	mc1	42.59	±	1.73
		mc2	58.44	±	1.58

Average: 58.44%

Bigbench

Task	Version	Metric	Value		Stderr
bigbench_causal_judgement	0	multiple_choice_grade	55.26	±	3.62
bigbench_date_understanding	0	multiple_choice_grade	64.77	±	2.49
bigbench_disambiguation_qa	0	multiple_choice_grade	37.60	±	3.02
bigbench_geometric_shapes	0	multiple_choice_grade	32.31	±	2.47
		exact_str_match	21.45	±	2.17
bigbench_logical_deduction_five_objects	0	multiple_choice_grade	31.00	±	2.07
bigbench_logical_deduction_seven_objects	0	multiple_choice_grade	22.43	±	1.58
bigbench_logical_deduction_three_objects	0	multiple_choice_grade	53.00	±	2.89
bigbench_movie_recommendation	0	multiple_choice_grade	40.40	±	2.20
bigbench_navigate	0	multiple_choice_grade	51.30	±	1.58
bigbench_reasoning_about_colored_objects	0	multiple_choice_grade	68.50	±	1.04
bigbench_ruin_names	0	multiple_choice_grade	48.66	±	2.36
bigbench_salient_translation_error_detection	0	multiple_choice_grade	30.36	±	1.46
bigbench_snarks	0	multiple_choice_grade	70.17	±	3.41
bigbench_sports_understanding	0	multiple_choice_grade	70.39	±	1.45
bigbench_temporal_sequences	0	multiple_choice_grade	31.00	±	1.46
bigbench_tracking_shuffled_objects_five_objects	0	multiple_choice_grade	21.44	±	1.16
bigbench_tracking_shuffled_objects_seven_objects	0	multiple_choice_grade	18.29	±	0.92
bigbench_tracking_shuffled_objects_three_objects	0	multiple_choice_grade	53.00	±	2.89

Average: 44.44%

Average score: 54.95%

Elapsed time: 02:50:51

🧩 Configuration

models:
  - model: mayflowergmbh/Wiedervereinigung-7b-dpo-laser
    # no parameters necessary for base model
  - model: flemmingmiguel/NeuDist-Ro-7B
    parameters:
      density: 0.60
      weight: 0.30
  - model: Blizado/discolm-mfto-7b-german-v0.1
    parameters:
      density: 0.65
      weight: 0.40
  - model: ResplendentAI/Flora_DPO_7B
    parameters:
      density: 0.6
      weight: 0.3
merge_method: dare_ties
base_model: mayflowergmbh/Wiedervereinigung-7b-dpo-laser
parameters:
  int8_mask: true
dtype: bfloat16
random_seed: 0
tokenizer_source: base

💻 Usage

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "cstr/Spaetzle-v12-7b"
messages = [{"role": "user", "content": "What is a large language model?"}]

tokenizer = AutoTokenizer.from_pretrained(model)
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    torch_dtype=torch.float16,
    device_map="auto",
)

outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])