metadata

license: apache-2.0
datasets:
  - argilla/distilabel-intel-orca-dpo-pairs
base_model: teknium/OpenHermes-2.5-Mistral-7B
model-index:
  - name: kellemar-DPO-7B-v1.01
    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: 65.78
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=decruz07/kellemar-DPO-7B-v1.01
          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: 85.04
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=decruz07/kellemar-DPO-7B-v1.01
          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: 63.24
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=decruz07/kellemar-DPO-7B-v1.01
          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: 55.54
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=decruz07/kellemar-DPO-7B-v1.01
          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: 78.69
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=decruz07/kellemar-DPO-7B-v1.01
          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: 61.64
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=decruz07/kellemar-DPO-7B-v1.01
          name: Open LLM Leaderboard

Model Card for decruz07/kellemar-DPO-7B-v1.01

This model was created using OpenHermes-2.5 as the base, and finetuned with argilla/distilabel-intel-orca-dpo-pairs.

Model Details

Finetuned with these specific parameters: Steps: 200 Learning Rate: 5e5 Beta: 0.1

Model Description

Developed by: @decruz
Funded by [optional]: my full-time job
Finetuned from model [optional]: teknium/OpenHermes-2.5-Mistral-7B

Benchmarks

OpenLLM

Average	ARC	HellaSwag	MMLU	TruthfulQA	Winogrande	GSM8K
68.32	65.78	85.04	63.24	55.54	78.69	61.64

Nous

AGIEval	GPT4All	TruthfulQA	Bigbench	Average
43.17	73.25	55.87	42.2	53.62

Uses

You can use this for basic inference. You could probably finetune with this if you want to.

How to Get Started with the Model

You can create a space out of this, or use basic python code to call the model directly and make inferences to it.

[More Information Needed]

Training Details

The following was used: `training_args = TrainingArguments( per_device_train_batch_size=4, gradient_accumulation_steps=4, gradient_checkpointing=True, learning_rate=5e-5, lr_scheduler_type="cosine", max_steps=200, save_strategy="no", logging_steps=1, output_dir=new_model, optim="paged_adamw_32bit", warmup_steps=100, bf16=True, report_to="wandb", )

Create DPO trainer

dpo_trainer = DPOTrainer( model, ref_model, args=training_args, train_dataset=dataset, tokenizer=tokenizer, peft_config=peft_config, beta=0.1, max_prompt_length=1024, max_length=1536, )`

Training Data

This was trained with https://huggingface.co/datasets/argilla/distilabel-intel-orca-dpo-pairs

Training Procedure

Trained with Labonne's Google Colab Notebook on Finetuning Mistral 7B with DPO.

Model Card Authors [optional]

@decruz

Model Card Contact

@decruz on X/Twitter

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	68.32
AI2 Reasoning Challenge (25-Shot)	65.78
HellaSwag (10-Shot)	85.04
MMLU (5-Shot)	63.24
TruthfulQA (0-shot)	55.54
Winogrande (5-shot)	78.69
GSM8k (5-shot)	61.64