Edit model card

thumbnail

djinn

djinn is a merge of the following models using LazyMergekit:

πŸ† Benchmarks

Nous benchmarks, find more details here

Model AGIEval GPT4All TruthfulQA Bigbench Average
chatty-djinn-14B 38.43 76.29 68.02 47.6 57.59

AGIEval

Task Version Metric Value Stderr
agieval_aqua_rat 0 acc 23.62 Β± 2.67
acc_norm 21.65 Β± 2.59
agieval_logiqa_en 0 acc 32.26 Β± 1.83
acc_norm 33.79 Β± 1.86
agieval_lsat_ar 0 acc 23.04 Β± 2.78
acc_norm 23.04 Β± 2.78
agieval_lsat_lr 0 acc 38.82 Β± 2.16
acc_norm 39.22 Β± 2.16
agieval_lsat_rc 0 acc 59.48 Β± 3.00
acc_norm 54.65 Β± 3.04
agieval_sat_en 0 acc 75.73 Β± 2.99
acc_norm 74.27 Β± 3.05
agieval_sat_en_without_passage 0 acc 35.92 Β± 3.35
acc_norm 34.47 Β± 3.32
agieval_sat_math 0 acc 31.36 Β± 3.14
acc_norm 26.36 Β± 2.98

Average: 38.43%

GPT4All

Task Version Metric Value Stderr
arc_challenge 0 acc 62.12 Β± 1.42
acc_norm 65.44 Β± 1.39
arc_easy 0 acc 83.88 Β± 0.75
acc_norm 78.58 Β± 0.84
boolq 1 acc 88.07 Β± 0.57
hellaswag 0 acc 65.18 Β± 0.48
acc_norm 86.45 Β± 0.34
openbookqa 0 acc 39.60 Β± 2.19
acc_norm 48.60 Β± 2.24
piqa 0 acc 82.26 Β± 0.89
acc_norm 83.62 Β± 0.86
winogrande 0 acc 83.27 Β± 1.05

Average: 76.29%

TruthfulQA

Task Version Metric Value Stderr
truthfulqa_mc 1 mc1 50.55 Β± 1.75
mc2 68.02 Β± 1.52

Average: 68.02%

Bigbench

Task Version Metric Value Stderr
bigbench_causal_judgement 0 multiple_choice_grade 57.89 Β± 3.59
bigbench_date_understanding 0 multiple_choice_grade 64.50 Β± 2.49
bigbench_disambiguation_qa 0 multiple_choice_grade 32.56 Β± 2.92
bigbench_geometric_shapes 0 multiple_choice_grade 26.18 Β± 2.32
exact_str_match 1.11 Β± 0.55
bigbench_logical_deduction_five_objects 0 multiple_choice_grade 30.80 Β± 2.07
bigbench_logical_deduction_seven_objects 0 multiple_choice_grade 22.86 Β± 1.59
bigbench_logical_deduction_three_objects 0 multiple_choice_grade 57.67 Β± 2.86
bigbench_movie_recommendation 0 multiple_choice_grade 62.00 Β± 2.17
bigbench_navigate 0 multiple_choice_grade 56.20 Β± 1.57
bigbench_reasoning_about_colored_objects 0 multiple_choice_grade 65.65 Β± 1.06
bigbench_ruin_names 0 multiple_choice_grade 64.73 Β± 2.26
bigbench_salient_translation_error_detection 0 multiple_choice_grade 17.33 Β± 1.20
bigbench_snarks 0 multiple_choice_grade 76.24 Β± 3.17
bigbench_sports_understanding 0 multiple_choice_grade 75.15 Β± 1.38
bigbench_temporal_sequences 0 multiple_choice_grade 48.90 Β± 1.58
bigbench_tracking_shuffled_objects_five_objects 0 multiple_choice_grade 22.32 Β± 1.18
bigbench_tracking_shuffled_objects_seven_objects 0 multiple_choice_grade 18.17 Β± 0.92
bigbench_tracking_shuffled_objects_three_objects 0 multiple_choice_grade 57.67 Β± 2.86

Average: 47.6%

Average score: 57.59%

🧩 Configuration

Inspired by theprofessor's config

merge_method: linear # use linear so we can include multiple models, albeit at a zero weight
parameters:
  weight: 1.0 # weight everything as 1 unless specified otherwise - linear with one model weighted at 1 is a no-op like passthrough
slices:
  - sources:
      - model: openchat/openchat-3.5-0106
        layer_range: [0, 1]
      - model: teknium/OpenHermes-2.5-Mistral-7B 
        layer_range: [0, 1]
        parameters:
          weight: 0
  - sources:
      - model: bardsai/jaskier-7b-dpo-v6.1
        layer_range: [1, 10]
  - sources:
      - model: senseable/WestLake-7B-v2
        layer_range: [10, 20]
  - sources:
      - model: NousResearch/Nous-Hermes-2-Mistral-7B-DPO
        layer_range: [20, 30]
  - sources:
      - model: paulml/OGNO-7B
        layer_range: [15, 25]
  - sources:
      - model: paulml/DPOB-INMTOB-7B
        layer_range: [22, 32]
  - sources:
      - model: mlabonne/AlphaMonarch-7B
        layer_range: [5, 15]
  - sources: 
      - model: openchat/openchat-3.5-0106
        layer_range: [31, 32]
      - model: teknium/OpenHermes-2.5-Mistral-7B
        layer_range: [31, 32]
        parameters:
          weight: 0
dtype: float16
tokenizer_source: model:openchat/openchat-3.5-0106

An example

User: Sally (a girl) has 3 brothers. Each brother has 2 sisters. How many sisters does Sally have? Think step by step and the result MUST BE A NUMBER

djinn: In this family structure, we know there are 4 aka '1 + 3-brothers' (1 being initially for Sally) existing individuals who are also the siblings of the 3-brother-group. As each of the 3-brothers has 2-sisters each, that sums up to 6 total sister- roles. Since we already accounted for one of the sisters (assuming to be the oldest or in alphabetical order, as names were not given) to be part of the 3-brother group, we subtract one from the 6-sisters total. So, considering the provided information, Sally would have 6 - 1 = 5 remaining younger or alphabetically-wise- placed, biological-sisters. However, note that this problem statement might not perfectly align with real-life complexities, such as same-sex twins, adoption, etc., but within the given context, the solution is 5.

Remember, in more intricate or unclear family tree scenarios, it'd be best to directly ask for clarifications or consider additional data points. In short, with the given constraints, Sally has 5 known, direct, biological, and likely-in-age-order, younger/alphabetically-placed, sister-siblings.(To reiterate the core result in a concise manner: 5).

(Please excuse the verbosity, aiming for thoroughness and clarity.)

This response maintains the initial request's numerical focus while providing a detailed explanation to ensure understanding.

πŸ’» Usage

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "mayacinka/djinn"
messages = [{"role": "user", "content": "What is a large language model?"}]

tokenizer = AutoTokenizer.from_pretrained(model)
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    torch_dtype=torch.float16,
    device_map="auto",
)

outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric Value
Avg. 72.08
AI2 Reasoning Challenge (25-Shot) 70.39
HellaSwag (10-Shot) 86.45
MMLU (5-Shot) 64.40
TruthfulQA (0-shot) 67.57
Winogrande (5-shot) 83.11
GSM8k (5-shot) 60.58
Downloads last month
907
Safetensors
Model size
13.6B params
Tensor type
FP16
Β·
Inference API
Input a message to start chatting with mayacinka/chatty-djinn-14B.
Model is too large to load in Inference API (serverless). To try the model, launch it on Inference Endpoints (dedicated) instead.

Merge of

Evaluation results