serhii.net

In the middle of the desert you can say anything you want

29 Jan 2024

Writing evaluation code for my Masterarbeit

Previously:

As before, lmentry code is a big inspiration.

Additionally:

I didn’t want to write an eval harness, but somehow I find myself doing that — but instead of a benchmark thing, as one-time task, and worse than the existing ones. I wonder.

Again walking through existing evals

OpenAI evals

openai/evals: Evals is a framework for evaluating LLMs and LLM systems, and an open-source registry of benchmarks.

  • evals/docs/build-eval.md at main · openai/evals
    • Each JSON object will represent one data point in your eval. The keys you need in the JSON object depend on the eval template. All templates expect an “input” key, which is the prompt, ideally specified in chat format (though strings are also supported). We recommend chat format even if you are evaluating non-chat models. If you are evaluating both chat and non-chat models, we handle the conversion between chat-formatted prompts and raw string prompts (see the conversion logic here).

      • Do I have any reasons for not exporting my code bits to a jsonl file with standard keys?
  • Example of an eval: evals/evals/registry/data/README.md at main · openai/evals
    • Input in Chat format
    • I love how ideal is a list of options, like [11, "11"].
  • Many non-English evals! EVEN UKRAINIAN ONES evals/evals/registry/data at main · openai/evals
    {"input": [{"role": "system", "content": "Ви отримаєте текст електронної петиції. Вам потрібно проаналізувати суть звернення та опираючись на законодавчу базу України та інші фактори відповісти чи підтримали би уряд цю петицію. Поясніть свій хід думок та висновок з позиції законодавства України."}, {"role": "user", "content": "Суть звернення: Повернути пільги на оплату електроенергії для населення, яке проживає у 30-кілометровій зоні атомних електростанцій.  Відновити інші пільги населенню на оплату спожитої електричної енергії. Дата складання петиції - 2021 рік."}], "ideal": "Уряд не підтримав цю петицію, оскільки вважає, що питання надання пільг та субсидій на оплату комунальних послуг, в тому числі електроенергії, є повноваженням Кабінету Міністрів України а не уряду. Крім того, уряд вважає, що в державному бюджеті України на 2021 рік вже передбачено достатній обсяг коштів для компенсації витрат вразливим верствам населення, у тому числі для населення, що проживає в 30-кілометровій зоні атомних електростанцій."}
    
  • YAML with LMs, exact names and metadata for them: evals/evals/registry/completion_fns/langchain_llms.yaml at main · openai/evals

OK I’m definitely doing that.

And the example/parsing bit is important, since by default it’s often more verbose than I’d like: 2024-01-29-213023_629x278_scrot.png

EleutherAI Evaluation harness

Desiderata/TODOs for my case

Looking at the above:

  • Main question: OpenAI Chat completion API VS Eleuther classic thing? + How do I integrate both?
  • My datasets will live on HF hub, more or less consistent in their column names
  • Datasets are a separate thing from what gets ‘fed’ to the eval
    • I generate that during eval through templates?

SO:

  • => Include semi-natively chat-completion-style instructions to my dataset dataclasses?

  • I love EleutherAI and Zeno and will be mainly using that! Instead of writing my own NIH bad eval package

  • Make all generators create dataclass-wizard-jsons AND flattened CSVs for all the tasks

  • CSV->HF in the eval package, together with the yamls for config

  • Oh look cbt · Datasets at Hugging Face

New eval-ua-tion package concept

  • It will have:
    • In: CSV? JSONs? w/ the dataset, that it will convert to HF and whatever
    • It will have the yaml for tasks descriptions of the tasks to feed eval-lm
    • it will have the eval-lm package, as well as the logic to run it (Dockerfile / Rancher pod YAML / ..) and save ti (??? as of yet)
    • It may have some bits for analyzing/plotting the evaluation results

Relevant

Interesting models

  • HF
    • mistralai/Mistral-7B-Instruct-v0.2
      • didn’t have enough patience to wait for one instance
    • TinyLlama/TinyLlama-1.1B-Chat-v1.0
      • easy to run on CPU for testing

Running stuff

Created a docker w/ lm-eval, interactively playing with it

  • cool params
    • --limit 1
    • --device=cpu is a thing

Was able to run this on CPU!

root@88265fe7e6e4:/lm-evaluation-harness
python3 -m lm_eval --model hf --model_args pretrained=TinyLlama/TinyLlama-1.1B-Chat-v1.0 --limit 1 --write_out --log_samples --output_path /tmp/outpt --tasks truthfulqa --device cpu

Generated this, took 19 minutes

: None, batch_size: 1
|      Tasks      |Version|Filter|n-shot|  Metric   | Value |   |Stderr|
|-----------------|-------|------|-----:|-----------|------:|---|------|
|truthfulqa       |N/A    |none  |     0|acc        | 0.9251|±  |N/A   |
|                 |       |none  |     0|bleu_max   | 8.9138|±  |N/A   |
|                 |       |none  |     0|bleu_acc   | 0.0000|±  |N/A   |
|                 |       |none  |     0|bleu_diff  | 0.0000|±  |N/A   |
|                 |       |none  |     0|rouge1_max |46.1538|±  |N/A   |
|                 |       |none  |     0|rouge1_acc | 1.0000|±  |N/A   |
|                 |       |none  |     0|rouge1_diff| 3.2967|±  |N/A   |
|                 |       |none  |     0|rouge2_max |18.1818|±  |N/A   |
|                 |       |none  |     0|rouge2_acc | 1.0000|±  |N/A   |
|                 |       |none  |     0|rouge2_diff| 1.5152|±  |N/A   |
|                 |       |none  |     0|rougeL_max |46.1538|±  |N/A   |
|                 |       |none  |     0|rougeL_acc | 1.0000|±  |N/A   |
|                 |       |none  |     0|rougeL_diff| 3.2967|±  |N/A   |
| - truthfulqa_gen|      3|none  |     0|bleu_max   | 8.9138|±  |N/A   |
|                 |       |none  |     0|bleu_acc   | 0.0000|±  |N/A   |
|                 |       |none  |     0|bleu_diff  | 0.0000|±  |N/A   |
|                 |       |none  |     0|rouge1_max |46.1538|±  |N/A   |
|                 |       |none  |     0|rouge1_acc | 1.0000|±  |N/A   |
|                 |       |none  |     0|rouge1_diff| 3.2967|±  |N/A   |
|                 |       |none  |     0|rouge2_max |18.1818|±  |N/A   |
|                 |       |none  |     0|rouge2_acc | 1.0000|±  |N/A   |
|                 |       |none  |     0|rouge2_diff| 1.5152|±  |N/A   |
|                 |       |none  |     0|rougeL_max |46.1538|±  |N/A   |
|                 |       |none  |     0|rougeL_acc | 1.0000|±  |N/A   |
|                 |       |none  |     0|rougeL_diff| 3.2967|±  |N/A   |
| - truthfulqa_mc1|      2|none  |     0|acc        | 1.0000|±  |N/A   |
| - truthfulqa_mc2|      2|none  |     0|acc        | 0.7752|±  |N/A   |

|  Groups  |Version|Filter|n-shot|  Metric   | Value |   |Stderr|
|----------|-------|------|-----:|-----------|------:|---|------|
|truthfulqa|N/A    |none  |     0|acc        | 0.9251|±  |N/A   |
|          |       |none  |     0|bleu_max   | 8.9138|±  |N/A   |
|          |       |none  |     0|bleu_acc   | 0.0000|±  |N/A   |
|          |       |none  |     0|bleu_diff  | 0.0000|±  |N/A   |
|          |       |none  |     0|rouge1_max |46.1538|±  |N/A   |
|          |       |none  |     0|rouge1_acc | 1.0000|±  |N/A   |
|          |       |none  |     0|rouge1_diff| 3.2967|±  |N/A   |
|          |       |none  |     0|rouge2_max |18.1818|±  |N/A   |
|          |       |none  |     0|rouge2_acc | 1.0000|±  |N/A   |
|          |       |none  |     0|rouge2_diff| 1.5152|±  |N/A   |
|          |       |none  |     0|rougeL_max |46.1538|±  |N/A   |
|          |       |none  |     0|rougeL_acc | 1.0000|±  |N/A   |
|          |       |none  |     0|rougeL_diff| 3.2967|±  |N/A   |
pretrained__TinyLlama__TinyLlama-1.1B-Chat-v1.0_truthfulqa_gen.jsonl
pretrained__TinyLlama__TinyLlama-1.1B-Chat-v1.0_truthfulqa_mc1.jsonl
pretrained__TinyLlama__TinyLlama-1.1B-Chat-v1.0_truthfulqa_mc2.jsonl
results.json

results contains a lot, the other files contain the exact document IDs, the used prompts, etc. — perfect, it works!Go

Game plan

  • I’ll try to avoid having installed the 5gb dependencies of lm-eval in the project

  • They will be installed in the Docker image

  • The project will contain only the yamls for my tasks

    • They will be included with --include_path in the runner
      • Tried it, it works!
    • You can allegedly also directly pass a yaml path to --tasks
  • Unsolved

    • Where to save results?
    • Rancher space thing, whatever it’s called?
    • scp them somewhere?

First custom task

Had a dataset on HF, used it:

task: pravda
dataset_path: shamotskyi/ukr_pravda_2y
dataset_name: null
# output_type: multiple_choice
training_split: null
validation_split: null
test_split: train
doc_to_text: "Predict a title for the following news: {{eng_text}}"
doc_to_target: "{{eng_title}}"
# doc_to_choice: "{{choices.text}}"
# should_decontaminate: true
# doc_to_decontamination_query: question_stem
metric_list:
  - metric: bleu
    aggregation: mean
    higher_is_better: true
metadata:
  version: 1.0

Changed metric to bleu, and used my rows.

Problem: some of the rows are null for the English text.

datasets.exceptions.DatasetGenerationCastError: An error occurred while generating the dataset

All the data files must have the same columns, but at some point there are 6 new columns (id, lang, kind, uri, date, domain) and 20 missing columns (rus_title, eng_text, tags, ukr_tags_full, rus_uri, rus_tags, ukr_text, date_published, eng_tags, rus_text, eng_title, ukr_author_name, ukr_uri, eng_uri, eng_tags_full, ukr_title, rus_author_name, eng_author_name, rus_tags_full, ukr_tags).

OK then :( all have to be equal

Using a local dataset

Local dataset or model path support · Issue #1224 · EleutherAI/lm-evaluation-harness showed how to use a local HF dataset (not json as shown in the tutorial):

task: lmentry
dataset_path: arrow
dataset_kwargs:
  data_files:
    train: /resources/ds/dataset/hf_WordsAlphabetOrder/data-00000-of-00001.arrow
# dataset_name: null
# output_type: multiple_choice
training_split: null
validation_split: null
test_split: train
doc_to_text: "{{question}}"
doc_to_target: "{{correctAnswer}}"
metric_list:
  - metric: bleu
#    aggregation: mean
#    higher_is_better: true
metadata:
version: 1.0

THIS GAVE ME THE FIRST NON-1.0 SCORE! I just had to use more test instances

root@lm-eval-sh:/lm-evaluation-harness# python3 -m lm_eval --model hf --model_args pretrained=TinyLlama/TinyLlama-1.1B-Chat-v1.0 --limit 520 --write_out --log_samples --output_path /tmp/Output --tasks lmentry --include_path /resources  --verbosity DEBUG --show_config

okay!

hf (pretrained=mistralai/Mistral-7B-Instruct-v0.2), gen_kwargs: (None), limit: 20000.0, num_fewshot: None, batch_size: 1
| Tasks |Version|Filter|n-shot|Metric|Value|   |Stderr|
|-------|------:|------|-----:|------|----:|---|-----:|
|lmentry|      1|none  |     0|acc   |0.485|±  |0.0354|

hf (pretrained=mistralai/Mistral-7B-Instruct-v0.2), gen_kwargs: (None), limit: 20000.0, num_fewshot: 2, batch_size: 1
| Tasks |Version|Filter|n-shot|Metric|Value|   |Stderr|
|-------|------:|------|-----:|------|----:|---|-----:|
|lmentry|      1|none  |     2|acc   |0.685|±  |0.0329|

hf (pretrained=mistralai/Mistral-7B-Instruct-v0.2), gen_kwargs: (None), limit: 20000.0, num_fewshot: 10, batch_size: 1
| Tasks |Version|Filter|n-shot|Metric|Value|   |Stderr|
|-------|------:|------|-----:|------|----:|---|-----:|
|lmentry|      1|none  |    10|acc   | 0.78|±  |0.0294|

OK! Increasing num_fewshot on that exact same test set predictably increases scores. OK, it all starts to make sense <3

So, fazit:

  • accuracy version breaks
  • multi-choice one works more or less predictably, but <0.5 with zero-shot?

Either way goal was to run an eval that at least runs, mission accomplished.

Onwards

non-English multichoice example:

I now understand why non-mc tasks failed with acc metric.

task: lmentry_low
dataset_path: arrow
dataset_kwargs:
  data_files:
    train: /datasets/hf_LOWTask/data-00000-of-00001.arrow
# dataset_name: null
#output_type: multiple_choice
training_split: null
validation_split: null
test_split: train
doc_to_text: "{{question}}"
doc_to_target: "{{correctAnswer}}"
#doc_to_choice: "{{[additionalMetadata_option_0, additionalMetadata_option_1]}}"
# doc_to_choice: "{{['yes', 'no']}}"
# should_decontaminate: true
# doc_to_decontamination_query: question_stem
metric_list:
  - metric: exact_match
    aggregation: mean
    higher_is_better: true
    ignore_case: true
    ignore_punctuation: true
metadata:
  version: 1.0
python3 -m lm_eval \
--model hf \
--model_args pretrained=mistralai/Mistral-7B-v0.1 \
--limit 100 \
--write_out \
--log_samples \
--output_path /MOutput \
--tasks low \
--device cuda \
--verbosity DEBUG \
--include_path /resources \
--show_config \
--num_fewshot 2

Useful bits for tasks

lm-evaluation-harness/lm_eval/tasks/unscramble/cycle_letters.yaml at big-refactor · EleutherAI/lm-evaluation-harness:

metric_list:
  - metric: exact_match
    aggregation: mean
    higher_is_better: true
    ignore_case: true
    ignore_punctuation: true

I can do doc_to_text: "{{system_prompts[0]}}. {{question}}"

Knowing when to stop
 "arguments": [
      [
        "Ви розв'язуєте екзамен з української мови. Вкажіть правильну відповідь одним словом, без лапок. Наприклад: \\n Питання: В слові \"герметизація\" яка літера третя?\\n Відповідь: р. Яка літера в слові \"собака\" перша?",
        {
          "until": [
            "\n\n"
          ],
          "do_sample": false
        }
      ]
    ],
    "resps": [
      [
        "\\n Відповідь: с. Яка літера в слові \"політика\" четверта?\\n Відповідь: т. Яка літера в слові \"політика\" п'ята?\\n Відповідь: к. Яка літера в слові \"політика\" шоста?\\n Відповідь: і. Яка літера в слові \"політика\" сьома?\\n Відповідь: т. Яка літера в слові \"політика\" восьма?\\n Відповідь: к. Яка літера в слові \"політика\" дев'ята?\\n Відповідь: а. Яка літера в слові \"політика\" десята?\\n Відповідь: л. Яка літера в слові \"політика\" одинадцята?\\n Відповідь: і. Яка літера в слові \"політика\" дванадцята?\\n Відпов"
      ]
    ],

is important it seems, haha. And editing my own examples is important as well if I manually inject system prompts instead of n_shot:

    "target": "с",
    "arguments": [
      [
        "Ви розв'язуєте екзамен з української мови. Вкажіть правильну відповідь одним словом, без лапок. Наприклад: \\n Питання: В слові \"герметизація\" яка літера третя?\\n Відповідь: р. В слові \"собака\" на першому місці знаходиться літера ...",
        {
          "until": [
            "\n\n"
          ],
          "do_sample": false
        }
      ]
output_type: generate_until
target_delimiter: ""
generation_kwargs:
  until:
    - "\n\n"
    - "\n"
  do_sample: false
  temperature: 0.0
target_delimiter: " "
metric_list:
  - metric: exact_match
    aggregation: mean
    higher_is_better: true
    ignore_case: true
    ignore_punctuation: true
filter_list:
  - name: "get-answer"
    filter:
      - function: "regex"
        regex_pattern: "The answer is (\\-?[0-9\\.\\,]+)"
      - function: "take_first"
filter_list:
  - name: remove_whitespace
    filter:
      - function: remove_whitespace
      - function: take_first

(from mgsm/en_cot/cot_yaml)

ag generation -A 8 helps find examples

I can’t find any good documentation on many of the params used.

  • About the results of WizardMath on GSM8K · Issue #1274 · EleutherAI/lm-evaluation-harness
    • For the base gsm8k task, we match the format used by the original GSM8k publication, where the format is Q: <question> \nA: <reasoning chain> #### <numeric answer> and are strict about only extracting an answer from the format #### <numeric answer>. Because models don’t know to output this format, they do not perform well 0-shot on it, but can do so few-shot.

So many things to learn from issues instead of documentation: always get acc,acc_norm, perplexity =1 on triviaqa task based on llama2 model · Issue 1239 · EleutherAI/lm-evaluation-harness

This worldlengthcomparison task gets a whopping 0.62 w/ mistral7b-notistruct using the same formulation as the others:

task: wlc_nomulti
group: lmentry
dataset_path: arrow
dataset_kwargs:
  data_files:
    train: /datasets/hf_WordLengthComparison/train/data-00000-of-00001.arrow
    test: /datasets/hf_WordLengthComparison/test/data-00000-of-00001.arrow
# dataset_name: null
#output_type: generate_until
#num_fewshot: 3
generation_kwargs:
    until:
    - "\n\n"
    - "\n"
    - "."
#  max_length: 40
training_split: null
validation_split: null
test_split: train
fewshot_split: test
doc_to_text: "{{question}}"
doc_to_target: "{{correctAnswer}}"
#doc_to_choice: "{{[additionalMetadata_option_0, additionalMetadata_option_1]}}"
# doc_to_choice: "{{['yes', 'no']}}"
# should_decontaminate: true
# doc_to_decontamination_query: question_stem
metric_list:
  - metric: exact_match
    aggregation: mean
    higher_is_better: true
    ignore_case: true
    ignore_punctuation: true
metadata:
  version: 1.0
        starts = "(starts|begins)"

        base_patterns = [
            rf"The first letter is {answer}",
            rf"The first letter {of} {word} is {answer}",
            rf"{answer} is the first letter {of} {word}",
            rf"{word} {starts} with {answer}",
            rf"The letter that {word} {starts} with is {answer}",
            rf"{answer} is the starting letter {of} {word}",
            rf"{word}: {answer}",
            rf"First letter: {answer}",
        ]

Zeno

export ZENO_API_KEY=zen_xxxx

root@lm-eval-sh:/lm-evaluation-harness# pip install zeno-client==0.1.9

root@lm-eval-sh:/lm-evaluation-harness# PYTHONPATH=. python3 scripts/zeno_visualize.py  --data_path=/Output --project_name "test"

More edge cases

again, this would need to be filtered out. From prompts definitely, they need spaces. But also generate_until.

"arguments": [
  [
	"В слові \"їжа\" під номером один знаходиться літера ... ї\n\nВ слові \"синхрофазотрон\" під номером дев'ять знаходиться літера ...з\n\nЯка літера в слові \"ліжко\" перша? л\n\nЯка літера в слові \"їжа\" остання?",
	{
	  "until": [
		"\n\n"
	  ],
	  "do_sample": false
	}
  ]
],
"resps": [
  [
	"... я"
  ]
],
"filtered_resps": [
  "... я"
],
"bleu": [
  "а",
  "... я"
]

KRUK

robinhad/kruk: Ukrainian instruction-tuned language models and datasets oh damn

Nel mezzo del deserto posso dire tutto quello che voglio.