MarcMistral-7B is a merge of the following models using
LazyMergekit:
As an experiment to find the best base merge to further fine-tuning, expect a lot of experiments named using parts of the component models until a clear winner emerges in the benchmarks
In this case merging the highest MMLU merge with a high ARC merge to see which qualities remain untouched or improv
1slices:
2 - sources:
3 - model: nfaheem/Marcoroni-7b-DPO-Merge
4 layer_range: [0, 32]
5 - model: EmbeddedLLM/Mistral-7B-Merge-14-v0.5
6 layer_range: [0, 32]
7
8merge_method: slerp
9base_model: EmbeddedLLM/Mistral-7B-Merge-14-v0.5
10
11parameters:
12 t:
13 - filter: self_attn
14 value: [0, 0.5, 0.3, 0.7, 1]
15 - filter: mlp
16 value: [1, 0.5, 0.7, 0.3, 0]
17 - value: 0.5 # fallback for rest of tensors
18tokenizer_source: union
19
20dtype: bfloat16
1!pip install -qU transformers accelerate
2
3from transformers import AutoTokenizer
4import transformers
5import torch
6
7model = "flemmingmiguel/MarcMistral-7B"
8messages = [{"role": "user", "content": "What is a large language model?"}]
9
10tokenizer = AutoTokenizer.from_pretrained(model)
11prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
12pipeline = transformers.pipeline(
13 "text-generation",
14 model=model,
15 torch_dtype=torch.float16,
16 device_map="auto",
17)
18
19outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
20print(outputs[0]["generated_text"])
