MiniMerlin-3B / README.md
teilomillet's picture
Adding Evaluation Results
178f27a verified
|
raw
history blame
4.18 kB
metadata
language:
  - fr
  - en
license: apache-2.0
tags:
  - code
widget:
  - text: <s> [|User|] Comment faire un bon plat ? </s>[|Assistant|]
model-index:
  - name: MiniMerlin-3B
    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: 44.37
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=teilomillet/MiniMerlin-3B
          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: 66.56
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=teilomillet/MiniMerlin-3B
          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: 43.21
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=teilomillet/MiniMerlin-3B
          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: 47.07
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=teilomillet/MiniMerlin-3B
          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: 64.4
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=teilomillet/MiniMerlin-3B
          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: 20.17
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=teilomillet/MiniMerlin-3B
          name: Open LLM Leaderboard

SFT on a synthetic custom (french) dataset (2k), from general question answering, problem solving to code question. It's a POC.

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

model = AutoModelForCausalLM.from_pretrained(
    "teilomillet/MiniMerlin-3B",
    revision="0.1",
    return_dict=True,
    torch_dtype=torch.bfloat16,
    device_map='auto'
)

tokenizer = AutoTokenizer.from_pretrained("teilomillet/MiniMerlin-3B")
tokenizer.pad_token = tokenizer.eos_token

text = "[|User|] Comment faire un bon plat ? </s>[|Assistant|]"
inputs = tokenizer(text, return_tensors="pt").to(0)

outputs = model.generate(**inputs, max_new_tokens=800)
print(tokenizer.decode(outputs[0], skip_special_tokens=False))

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric Value
Avg. 47.63
AI2 Reasoning Challenge (25-Shot) 44.37
HellaSwag (10-Shot) 66.56
MMLU (5-Shot) 43.21
TruthfulQA (0-shot) 47.07
Winogrande (5-shot) 64.40
GSM8k (5-shot) 20.17