metadata

language:
  - de
  - en
license: cc-by-nc-4.0
tags:
  - merge
  - mergekit
  - lazymergekit
base_model:
  - abideen/AlphaMonarch-dora
  - mayflowergmbh/Wiedervereinigung-7b-dpo
  - flemmingmiguel/NeuDist-Ro-7B
  - ResplendentAI/Flora_DPO_7B
  - yleo/EmertonMonarch-7B
  - occiglot/occiglot-7b-de-en-instruct
  - OpenPipe/mistral-ft-optimized-1227
  - DiscoResearch/DiscoLM_German_7b_v1
  - LeoLM/leo-mistral-hessianai-7b
  - DRXD1000/Phoenix
  - VAGOsolutions/SauerkrautLM-7b-v1-mistral
  - malteos/hermeo-7b
  - FelixChao/WestSeverus-7B-DPO-v2
  - cognitivecomputations/openchat-3.5-0106-laser
model-index:
  - name: Spaetzle-v69-7b
    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: 69.54
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=cstr/Spaetzle-v69-7b
          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: 86.77
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=cstr/Spaetzle-v69-7b
          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: 64.63
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=cstr/Spaetzle-v69-7b
          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: 65.61
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=cstr/Spaetzle-v69-7b
          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: 81.93
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=cstr/Spaetzle-v69-7b
          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: 68.76
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=cstr/Spaetzle-v69-7b
          name: Open LLM Leaderboard

Spaetzle-v69-7b

This is a progressive (mostly dare-ties, but also slerp) merge with the intention of a suitable compromise for English and German local tasks.

There is also a 4q_k_m quantized GGUF.

It should work sufficiently well with ChatML prompt template (for all merged models should have seen ChatML prompts at least in DPO stage).

Evaluation

Benchmark scores are not the possible optimum, as the model attempts a compromise with a number of parameters, like German language performance, instruction following, reasoning capabilities, robustness (so far, i did not encounter inserted tokens, e.g.), model licensing, and other criteria. Nevertheless, they are not too bad:

It achieves (running quantized) in

German EQ Bench: Score (v2_de): 62.59 (Parseable: 171.0).
English EQ Bench: Score (v2): 76.43 (Parseable: 171.0).

Model	AGIEval	GPT4All	TruthfulQA	Bigbench	Average
Spaetzle-v69-7b	44.48	75.84	66.15	46.59	58.27

AGIEval

Task	Version	Metric	Value		Stderr
agieval_aqua_rat	0	acc	25.98	±	2.76
		acc_norm	23.62	±	2.67
agieval_logiqa_en	0	acc	39.78	±	1.92
		acc_norm	39.48	±	1.92
agieval_lsat_ar	0	acc	23.48	±	2.80
		acc_norm	23.91	±	2.82
agieval_lsat_lr	0	acc	50.00	±	2.22
		acc_norm	51.76	±	2.21
agieval_lsat_rc	0	acc	63.94	±	2.93
		acc_norm	64.31	±	2.93
agieval_sat_en	0	acc	76.70	±	2.95
		acc_norm	77.67	±	2.91
agieval_sat_en_without_passage	0	acc	46.12	±	3.48
		acc_norm	44.17	±	3.47
agieval_sat_math	0	acc	34.09	±	3.20
		acc_norm	30.91	±	3.12

Average: 44.48%

GPT4All

Task	Version	Metric	Value		Stderr
arc_challenge	0	acc	63.23	±	1.41
		acc_norm	64.16	±	1.40
arc_easy	0	acc	85.90	±	0.71
		acc_norm	82.49	±	0.78
boolq	1	acc	87.80	±	0.57
hellaswag	0	acc	67.05	±	0.47
		acc_norm	85.19	±	0.35
openbookqa	0	acc	38.40	±	2.18
		acc_norm	48.40	±	2.24
piqa	0	acc	82.75	±	0.88
		acc_norm	84.28	±	0.85
winogrande	0	acc	78.53	±	1.15

Average: 75.84%

TruthfulQA

Task	Version	Metric	Value		Stderr
truthfulqa_mc	1	mc1	50.67	±	1.75
		mc2	66.15	±	1.48

Average: 66.15%

Bigbench

Task	Version	Metric	Value		Stderr
bigbench_causal_judgement	0	multiple_choice_grade	56.84	±	3.60
bigbench_date_understanding	0	multiple_choice_grade	66.67	±	2.46
bigbench_disambiguation_qa	0	multiple_choice_grade	40.70	±	3.06
bigbench_geometric_shapes	0	multiple_choice_grade	24.79	±	2.28
		exact_str_match	10.58	±	1.63
bigbench_logical_deduction_five_objects	0	multiple_choice_grade	31.00	±	2.07
bigbench_logical_deduction_seven_objects	0	multiple_choice_grade	23.00	±	1.59
bigbench_logical_deduction_three_objects	0	multiple_choice_grade	58.00	±	2.85
bigbench_movie_recommendation	0	multiple_choice_grade	45.80	±	2.23
bigbench_navigate	0	multiple_choice_grade	52.10	±	1.58
bigbench_reasoning_about_colored_objects	0	multiple_choice_grade	69.55	±	1.03
bigbench_ruin_names	0	multiple_choice_grade	48.88	±	2.36
bigbench_salient_translation_error_detection	0	multiple_choice_grade	30.96	±	1.46
bigbench_snarks	0	multiple_choice_grade	73.48	±	3.29
bigbench_sports_understanding	0	multiple_choice_grade	74.14	±	1.40
bigbench_temporal_sequences	0	multiple_choice_grade	42.70	±	1.56
bigbench_tracking_shuffled_objects_five_objects	0	multiple_choice_grade	23.60	±	1.20
bigbench_tracking_shuffled_objects_seven_objects	0	multiple_choice_grade	18.40	±	0.93
bigbench_tracking_shuffled_objects_three_objects	0	multiple_choice_grade	58.00	±	2.85

Average: 46.59%

Average score: 58.27%

🧩 Merge Configuration

Spaetzle-v69-7b is a merge of the following models using LazyMergekit:

The merge tree in total involves the following original models:

For this last merge:

models:
  - model: cstr/Spaetzle-v68-7b
    # no parameters necessary for base model
  - model: abideen/AlphaMonarch-dora
    parameters:
      density: 0.60
      weight: 0.30
merge_method: dare_ties
base_model: cstr/Spaetzle-v68-7b
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-v69-7b"
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"])

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	72.87
AI2 Reasoning Challenge (25-Shot)	69.54
HellaSwag (10-Shot)	86.77
MMLU (5-Shot)	64.63
TruthfulQA (0-shot)	65.61
Winogrande (5-shot)	81.93
GSM8k (5-shot)	68.76