language:

• de license: apache-2.0 tags:
• hermeo
• laser datasets:
• LeoLM/OpenSchnabeltier pipeline_tag: conversational model-index:
• name: germeo-7b-laser 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: 60.75 name: normalized accuracy source: url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=aari1995/germeo-7b-laser 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: 82.81 name: normalized accuracy source: url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=aari1995/germeo-7b-laser 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: 60.57 name: accuracy source: url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=aari1995/germeo-7b-laser 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: 53.83 source: url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=aari1995/germeo-7b-laser 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: 75.61 name: accuracy source: url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=aari1995/germeo-7b-laser 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: 43.37 name: accuracy source: url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=aari1995/germeo-7b-laser name: Open LLM Leaderboard

(Evaluation WIP)

Hermes + Leo + German Laser = Germeo

Germeo-7B-Laser

A German-English understanding, but German-only speaking model merged from Hermeo-7B.

Model details

Merged from: leo-mistral-hessianai-7b-chat and DPOpenHermes-7B-v2

Model type: Causal decoder-only transformer language model

Languages: German replies with English Understanding Capabilities

Laser-Data: LeoLM/OpenSchnabeltier

This is an early experiment on laser and its influence on language understanding. It generally improves the language understanding capabilities. The hypothesis is that it degrades the probability of English replies and increasing those of German replies. The models internal German capabilities are boosted.

Will keep you updated..

Acknowledgements:

I would like to thank everyone that participated in making this model and its training possible: To @malteos for hermeo To @cognitivecomputations and Fernando Fernandes Neto for their implementation of LASER To @LeoLM and Björn for the OpenSchnabeltier dataset.

Prompt format:

streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
# Convert prompt to tokens
prompt_template = """<|im_start|>system
Du bist ein hilfreicher Assistent.<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant"""

prompt = "Schreibe eine Stellenanzeige für Data Scientist bei AXA!"

final_prompt = prompt_template.format(prompt=prompt)

Limit the model to output reply-only:

To solve this, you need to implement a custom stopping criteria:

from transformers import StoppingCriteria
class GermeoStoppingCriteria(StoppingCriteria):
  def __init__(self, target_sequence, prompt):
      self.target_sequence = target_sequence
      self.prompt=prompt

  def __call__(self, input_ids, scores, **kwargs):
      # Get the generated text as a string
      generated_text = tokenizer.decode(input_ids[0])
      generated_text = generated_text.replace(self.prompt,'')
      # Check if the target sequence appears in the generated text
      if self.target_sequence in generated_text:
          return True  # Stop generation

      return False  # Continue generation

  def __len__(self):
      return 1

  def __iter__(self):
      yield self

This then expects your input prompt (formatted as given into the model), and a stopping criteria, in this case the im_end token. Simply add it to the generation:

generation_output = model.generate(
    tokens, 
    streamer=streamer,
    max_new_tokens=1012,
    stopping_criteria=GermeoStoppingCriteria("<|im_end|>", prompt_template.format(prompt=prompt))
)

German benchmarks

Even though the model does not generate English text without being explicitly asked, performance on English Benchmarks is still up:

English benchmarks

Open LLM Leaderboard Evaluation Results

Detailed results can be found here