metadata

library_name: transformers
tags:
  - merge
  - llama
  - dpo
base_model:
  - akjindal53244/Llama-3.1-Storm-8B
  - Sao10K/L3.1-8B-Niitama-v1.1
  - v000000/L3.1-Niitorm-8B-t0.0001
datasets:
  - jondurbin/gutenberg-dpo-v0.1
model-index:
  - name: L3.1-Niitorm-8B-DPO-t0.0001
    results:
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: IFEval (0-Shot)
          type: HuggingFaceH4/ifeval
          args:
            num_few_shot: 0
        metrics:
          - type: inst_level_strict_acc and prompt_level_strict_acc
            value: 76.89
            name: strict accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=v000000/L3.1-Niitorm-8B-DPO-t0.0001
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: BBH (3-Shot)
          type: BBH
          args:
            num_few_shot: 3
        metrics:
          - type: acc_norm
            value: 30.51
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=v000000/L3.1-Niitorm-8B-DPO-t0.0001
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MATH Lvl 5 (4-Shot)
          type: hendrycks/competition_math
          args:
            num_few_shot: 4
        metrics:
          - type: exact_match
            value: 14.88
            name: exact match
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=v000000/L3.1-Niitorm-8B-DPO-t0.0001
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: GPQA (0-shot)
          type: Idavidrein/gpqa
          args:
            num_few_shot: 0
        metrics:
          - type: acc_norm
            value: 5.93
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=v000000/L3.1-Niitorm-8B-DPO-t0.0001
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MuSR (0-shot)
          type: TAUR-Lab/MuSR
          args:
            num_few_shot: 0
        metrics:
          - type: acc_norm
            value: 7.26
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=v000000/L3.1-Niitorm-8B-DPO-t0.0001
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MMLU-PRO (5-shot)
          type: TIGER-Lab/MMLU-Pro
          config: main
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 31.85
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=v000000/L3.1-Niitorm-8B-DPO-t0.0001
          name: Open LLM Leaderboard

Llama-3.1-Niitorm-8B-DPO

DPO Trained, Llama3.1-8B.

New: DPO'd Gutenberg Version (full epoch training).

RP model, Niitama 1.1 as a base, nearswapped with one of the smartest 3.1 models "Storm", then DPO'd, mostly abliterated.

Essentially, it's an improved Niitama 1.1

Gutenberg DPO creates more human-like prose/story writing and greately lessen synthetic feeling outputs.

llama.cpp:

thank you, mradermacher (GGUF)

thank you, QuantFactory (GGUF)

GGUF static

v0 (GGUF)

GGUF Imatrix -only q8, q6 k, q5 k s, q4 k s, iq4 x s

Finetune and merge

This is a merge and finetune of pre-trained language models.

Resultant merge finetuned on jondurbin/gutenberg-dpo-v0.1 for 1 epoch, 1.5e-5 learning rate, on Nvidia A100.

Merge Details

Merge Method

This model was merged using the NEARSWAP t0.0001 merge algorithm.

Models Merged

The following models were included in the merge:

Configuration

The following YAML configuration was used to produce this model:

slices:
  - sources:
      - model: Sao10K/L3.1-8B-Niitama-v1.1+grimjim/Llama-3-Instruct-abliteration-LoRA-8B
        layer_range: [0, 32]
      - model: akjindal53244/Llama-3.1-Storm-8B
        layer_range: [0, 32]
merge_method: nearswap
base_model: Sao10K/L3.1-8B-Niitama-v1.1+grimjim/Llama-3-Instruct-abliteration-LoRA-8B
parameters:
  t:
    - value: 0.0001
dtype: float16

# Then, DPO Finetune
# [jondurbin/gutenberg-dpo-v0.1](https://huggingface.co/datasets/jondurbin/gutenberg-dpo-v0.1)

DPO Notes

I used a higher learning rate and full dataset when training compared to my "L3.1-Celestial-Stone-2x8B-DPO". This caused lower loss and better adaption to the chosen style.

Prompt Template:

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

{system_prompt}<|eot_id|><|start_header_id|>user<|end_header_id|>

{input}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

{output}<|eot_id|>

Credit to Alchemonaut.

Credit to Sao10K.

Credit to Grimjim.

Credit to mlabonne.

Credit to jondurbin.

Credit to woofwolfy.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	27.89
IFEval (0-Shot)	76.89
BBH (3-Shot)	30.51
MATH Lvl 5 (4-Shot)	14.88
GPQA (0-shot)	5.93
MuSR (0-shot)	7.26
MMLU-PRO (5-shot)	31.85