aari1995/germeo-7b-laser

(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

German tasks:	MMLU-DE	Hellaswag-DE	ARC-DE	Average
Models / Few-shots:	(5 shots)	(10 shots)	(24 shots)
7B parameters
llama-2-7b	0.400	0.513	0.381	0.431
leo-hessianai-7b	0.400	0.609	0.429	0.479
bloom-6b4-clp-german	0.274	0.550	0.351	0.392
mistral-7b	0.524	0.588	0.473	0.528
leo-mistral-hessianai-7b	0.481	0.663	0.485	0.543
leo-mistral-hessianai-7b-chat	0.458	0.617	0.465	0.513
DPOpenHermes-7B-v2	0.517	0.603	0.515	0.545
hermeo-7b	0.511	0.668	0.528	0.569
germeo-7b-laser (this model)	?	?	?	?
13B parameters
llama-2-13b	0.469	0.581	0.468	0.506
leo-hessianai-13b	0.486	0.658	0.509	0.551
70B parameters
llama-2-70b	0.597	0.674	0.561	0.611
leo-hessianai-70b	0.653	0.721	0.600	0.658

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

English benchmarks

English tasks:	MMLU	Hellaswag	ARC	Average
Models / Few-shots:	(5 shots)	(10 shots)	(24 shots)
llama-2-7b	0.466	0.786	0.530	0.594
leolm-hessianai-7b	0.423	0.759	0.522	0.568
bloom-6b4-clp-german	0.264	0.525	0.328	0.372
mistral-7b	0.635	0.832	0.607	0.691
leolm-mistral-hessianai-7b	0.550	0.777	0.518	0.615
hermeo-7b	0.601	0.821	0.620	0.681
germeo-7b-laser (this model)	0.601	0.828	0.608	0.679

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	62.82
AI2 Reasoning Challenge (25-Shot)	60.75
HellaSwag (10-Shot)	82.81
MMLU (5-Shot)	60.57
TruthfulQA (0-shot)	53.83
Winogrande (5-shot)	75.61
GSM8k (5-shot)	43.37