mpasila's picture
Update README.md
21bfddc verified
metadata
language:
  - fi
license: apache-2.0
tags:
  - text-generation-inference
  - transformers
  - unsloth
  - llama
  - trl
  - sft
base_model: LumiOpen/Viking-7B
datasets:
  - mpasila/Finnish-Alpaca-Small

This is a merge of mpasila/Finnish-Alpaca-Small-LoRA-7B.

LoRA trained in 4-bit with 2k context using LumiOpen/Viking-7B as the base model for 1 epoch.

Dataset used is mpasila/Finnish-Alpaca-Small.

Re-trained because I have no idea if I used the fully trained model or the partially trained model (of Viking-7B), since it apparently was just released. (After re-training the score lowered noticeably so I wonder if I screwed up something.)

Prompt format: Alpaca

It uses Alpaca format but with a translated instruction at the start:

{
    "instruction,output": "Alla on ohje, jossa kuvataan tehtävä. Kirjoita vastaus, joka täyttää pyynnön asianmukaisesti.\n\n### Instruction:\n%instruction%\n\n### Response:\n%output%",
    "instruction,input,output": "Alla on ohje, jossa kuvataan tehtävä ja joka on yhdistetty kontekstia lisäävään syötteeseen. Kirjoita vastaus, joka täyttää pyynnön asianmukaisesti.\n\n### Instruction:\n%instruction%\n\n### Input:\n%input%\n\n### Response:\n%output%"
}

Evaluation

Model Size Type FIN-bench (score)
mpasila/Finnish-Alpaca-Small-7B 7B Instruct 0.3586
mpasila/Finnish-Alpaca-Tiny-V2-7B 7B Instruct 0.4654
mpasila/Alpacazord-Viking-7B 7B Instruct 0.4123
mpasila/NordicAlpaca-Finnish-V1-7B 7B Instruct 0.3891
mpasila/Finnish-Viking-Alpaca-V1-7B 7B Instruct 0.3943
Finnish-NLP/llama-7b-finnish-instruct-v0.1 7B Instruct 0.4365
Finnish-NLP/llama-7b-finnish-instruct-v0.2 7B Instruct 0.3993
Finnish-NLP/llama-7b-finnish 7B Base 0.2350
LumiOpen/Viking-7B (1000B) 7B Base 0.3721
HPLT/gpt-7b-nordic-prerelease 7B Base 0.3169

Source

FIN-bench scores:

Task Version Metric Value Stderr
bigbench_analogies 0 multiple_choice_grade 0.5923 ± 0.0433
bigbench_arithmetic_1_digit_addition 0 multiple_choice_grade 0.2700 ± 0.0446
bigbench_arithmetic_1_digit_division 0 multiple_choice_grade 0.4783 ± 0.1065
bigbench_arithmetic_1_digit_multiplication 0 multiple_choice_grade 0.2600 ± 0.0441
bigbench_arithmetic_1_digit_subtraction 0 multiple_choice_grade 0.2200 ± 0.0416
bigbench_arithmetic_2_digit_addition 0 multiple_choice_grade 0.1700 ± 0.0378
bigbench_arithmetic_2_digit_division 0 multiple_choice_grade 0.3600 ± 0.0482
bigbench_arithmetic_2_digit_multiplication 0 multiple_choice_grade 0.2000 ± 0.0402
bigbench_arithmetic_2_digit_subtraction 0 multiple_choice_grade 0.1300 ± 0.0338
bigbench_arithmetic_3_digit_addition 0 multiple_choice_grade 0.3100 ± 0.0465
bigbench_arithmetic_3_digit_division 0 multiple_choice_grade 0.2100 ± 0.0409
bigbench_arithmetic_3_digit_multiplication 0 multiple_choice_grade 0.1600 ± 0.0368
bigbench_arithmetic_3_digit_subtraction 0 multiple_choice_grade 0.2300 ± 0.0423
bigbench_arithmetic_4_digit_addition 0 multiple_choice_grade 0.3900 ± 0.0490
bigbench_arithmetic_4_digit_division 0 multiple_choice_grade 0.2300 ± 0.0423
bigbench_arithmetic_4_digit_multiplication 0 multiple_choice_grade 0.2100 ± 0.0409
bigbench_arithmetic_4_digit_subtraction 0 multiple_choice_grade 0.4500 ± 0.0500
bigbench_arithmetic_5_digit_addition 0 multiple_choice_grade 0.4800 ± 0.0502
bigbench_arithmetic_5_digit_division 0 multiple_choice_grade 0.0700 ± 0.0256
bigbench_arithmetic_5_digit_multiplication 0 multiple_choice_grade 0.1700 ± 0.0378
bigbench_arithmetic_5_digit_subtraction 0 multiple_choice_grade 0.5800 ± 0.0496
bigbench_cause_and_effect_one_sentence 0 multiple_choice_grade 0.6275 ± 0.0684
bigbench_cause_and_effect_one_sentence_no_prompt 0 multiple_choice_grade 0.6667 ± 0.0667
bigbench_cause_and_effect_two_sentences 0 multiple_choice_grade 0.5098 ± 0.0707
bigbench_emotions 0 multiple_choice_grade 0.3312 ± 0.0373
bigbench_empirical_judgments 0 multiple_choice_grade 0.3333 ± 0.0476
bigbench_general_knowledge 0 multiple_choice_grade 0.2857 ± 0.0544
bigbench_hhh_alignment_harmless 0 multiple_choice_grade 0.3793 ± 0.0643
bigbench_hhh_alignment_helpful 0 multiple_choice_grade 0.3559 ± 0.0629
bigbench_hhh_alignment_honest 0 multiple_choice_grade 0.3559 ± 0.0629
bigbench_hhh_alignment_other 0 multiple_choice_grade 0.5349 ± 0.0770
bigbench_intent_recognition 0 multiple_choice_grade 0.1546 ± 0.0138
bigbench_misconceptions 0 multiple_choice_grade 0.5448 ± 0.0432
bigbench_paraphrase 0 multiple_choice_grade 0.5300 ± 0.0354
bigbench_sentence_ambiguity 0 multiple_choice_grade 0.4333 ± 0.0645
bigbench_similarities_abstraction 0 multiple_choice_grade 0.6974 ± 0.0530

Uploaded model

  • Developed by: mpasila
  • License: apache-2.0
  • Finetuned from model : LumiOpen/Viking-7B

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.