Open Luodian opened 2 months ago
baberabb
commented
2 months ago I would look at the generations in the samples file, and also add some fewshots to the context (say --num_fewshot 5) to prompt the model with the desired format. Might have a bit more luck but pythia-160m is probably too small to be capable of cohesive generations.
Luodian
commented
2 months ago I think it's pretty weird, and it may not related to in-context learning. I also evaluate Qwen/Qwen2-0.5B, it's also 0-acc on mmlu_generative.
And I tested on mmlu_pro which is also a generative task, and it have normal accuracy.
hf (pretrained=Qwen/Qwen2-0.5B-Instruct), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 8
| Tasks |Version| Filter |n-shot| Metric | |Value | |Stderr|
|-------------------|------:|--------------|-----:|-----------|---|-----:|---|-----:|
|gpqa_main_zeroshot | 1|none | 0|acc |↑ |0.2857|± |0.0214|
| | |none | 0|acc_norm |↑ |0.2857|± |0.0214|
|mmlu_pro | 1|custom-extract| |exact_match|↑ |0.1444|± |0.0032|
| - biology | 0|custom-extract| 5|exact_match|↑ |0.2483|± |0.0161|
| - business | 0|custom-extract| 5|exact_match|↑ |0.1166|± |0.0114|
| - chemistry | 0|custom-extract| 5|exact_match|↑ |0.1025|± |0.0090|
| - computer_science| 0|custom-extract| 5|exact_match|↑ |0.1195|± |0.0160|
| - economics | 0|custom-extract| 5|exact_match|↑ |0.1979|± |0.0137|
| - engineering | 0|custom-extract| 5|exact_match|↑ |0.0918|± |0.0093|
| - health | 0|custom-extract| 5|exact_match|↑ |0.1467|± |0.0124|
| - history | 0|custom-extract| 5|exact_match|↑ |0.1706|± |0.0193|
| - law | 0|custom-extract| 5|exact_match|↑ |0.1317|± |0.0102|
| - math | 0|custom-extract| 5|exact_match|↑ |0.1288|± |0.0091|
| - other | 0|custom-extract| 5|exact_match|↑ |0.1591|± |0.0120|
| - philosophy | 0|custom-extract| 5|exact_match|↑ |0.1423|± |0.0157|
| - physics | 0|custom-extract| 5|exact_match|↑ |0.1101|± |0.0087|
| - psychology | 0|custom-extract| 5|exact_match|↑ |0.2268|± |0.0148|
| Groups |Version| Filter |n-shot| Metric | |Value | |Stderr|
|--------|------:|--------------|------|-----------|---|-----:|---|-----:|
|mmlu_pro| 1|custom-extract| |exact_match|↑ |0.1444|± |0.0032|Qwen2-0.5B-Instruct on mmlu_generative.
hf (pretrained=Qwen/Qwen2-0.5B-Instruct), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 8
| Tasks |Version|Filter|n-shot| Metric | |Value| |Stderr|
|---------------------------------------|------:|------|-----:|-----------|---|----:|---|-----:|
|mmlu (generative) | 2|none | |exact_match|↑ | 0|± | 0|
| - formal_logic | 2|none | 0|exact_match|↑ | 0|± | 0|
| - high_school_european_history | 2|none | 0|exact_match|↑ | 0|± | 0|
| - high_school_us_history | 2|none | 0|exact_match|↑ | 0|± | 0|
| - high_school_world_history | 2|none | 0|exact_match|↑ | 0|± | 0|
| - international_law | 2|none | 0|exact_match|↑ | 0|± | 0|
| - jurisprudence | 2|none | 0|exact_match|↑ | 0|± | 0|
| - logical_fallacies | 2|none | 0|exact_match|↑ | 0|± | 0|
| - moral_disputes | 2|none | 0|exact_match|↑ | 0|± | 0|
| - moral_scenarios | 2|none | 0|exact_match|↑ | 0|± | 0|
| - philosophy | 2|none | 0|exact_match|↑ | 0|± | 0|
| - prehistory | 2|none | 0|exact_match|↑ | 0|± | 0|
| - professional_law | 2|none | 0|exact_match|↑ | 0|± | 0|
| - world_religions | 2|none | 0|exact_match|↑ | 0|± | 0|
| - business_ethics | 2|none | 0|exact_match|↑ | 0|± | 0|
| - clinical_knowledge | 2|none | 0|exact_match|↑ | 0|± | 0|
| - college_medicine | 2|none | 0|exact_match|↑ | 0|± | 0|
| - global_facts | 2|none | 0|exact_match|↑ | 0|± | 0|
| - human_aging | 2|none | 0|exact_match|↑ | 0|± | 0|
| - management | 2|none | 0|exact_match|↑ | 0|± | 0|
| - marketing | 2|none | 0|exact_match|↑ | 0|± | 0|
| - medical_genetics | 2|none | 0|exact_match|↑ | 0|± | 0|
| - miscellaneous | 2|none | 0|exact_match|↑ | 0|± | 0|
| - nutrition | 2|none | 0|exact_match|↑ | 0|± | 0|
| - professional_accounting | 2|none | 0|exact_match|↑ | 0|± | 0|
| - professional_medicine | 2|none | 0|exact_match|↑ | 0|± | 0|
| - virology | 2|none | 0|exact_match|↑ | 0|± | 0|
| - econometrics | 2|none | 0|exact_match|↑ | 0|± | 0|
| - high_school_geography | 2|none | 0|exact_match|↑ | 0|± | 0|
| - high_school_government_and_politics| 2|none | 0|exact_match|↑ | 0|± | 0|
| - high_school_macroeconomics | 2|none | 0|exact_match|↑ | 0|± | 0|
| - high_school_microeconomics | 2|none | 0|exact_match|↑ | 0|± | 0|
| - high_school_psychology | 2|none | 0|exact_match|↑ | 0|± | 0|
| - human_sexuality | 2|none | 0|exact_match|↑ | 0|± | 0|
| - professional_psychology | 2|none | 0|exact_match|↑ | 0|± | 0|
| - public_relations | 2|none | 0|exact_match|↑ | 0|± | 0|
| - security_studies | 2|none | 0|exact_match|↑ | 0|± | 0|
| - sociology | 2|none | 0|exact_match|↑ | 0|± | 0|
| - us_foreign_policy | 2|none | 0|exact_match|↑ | 0|± | 0|
| - abstract_algebra | 2|none | 0|exact_match|↑ | 0|± | 0|
| - anatomy | 2|none | 0|exact_match|↑ | 0|± | 0|
| - astronomy | 2|none | 0|exact_match|↑ | 0|± | 0|
| - college_biology | 2|none | 0|exact_match|↑ | 0|± | 0|
| - college_chemistry | 2|none | 0|exact_match|↑ | 0|± | 0|
| - college_computer_science | 2|none | 0|exact_match|↑ | 0|± | 0|
| - college_mathematics | 2|none | 0|exact_match|↑ | 0|± | 0|
| - college_physics | 2|none | 0|exact_match|↑ | 0|± | 0|
| - computer_security | 2|none | 0|exact_match|↑ | 0|± | 0|
| - conceptual_physics | 2|none | 0|exact_match|↑ | 0|± | 0|
| - electrical_engineering | 2|none | 0|exact_match|↑ | 0|± | 0|
| - elementary_mathematics | 2|none | 0|exact_match|↑ | 0|± | 0|
| - high_school_biology | 2|none | 0|exact_match|↑ | 0|± | 0|
| - high_school_chemistry | 2|none | 0|exact_match|↑ | 0|± | 0|
| - high_school_computer_science | 2|none | 0|exact_match|↑ | 0|± | 0|
| - high_school_mathematics | 2|none | 0|exact_match|↑ | 0|± | 0|
| - high_school_physics | 2|none | 0|exact_match|↑ | 0|± | 0|
| - high_school_statistics | 2|none | 0|exact_match|↑ | 0|± | 0|
| - machine_learning | 2|none | 0|exact_match|↑ | 0|± | 0|
| Groups |Version|Filter|n-shot| Metric | |Value| |Stderr|
|-----------------|------:|------|------|-----------|---|----:|---|-----:|
|mmlu (generative)| 2|none | |exact_match|↑ | 0|± | 0|I'll take a look! My guess is a bug in the answer extraction
AishaAlaagib
commented
1 month ago Hello, I am having similar result (0 for all subtasks) and I am wondering if you have figured it out?
1436033631
commented
3 weeks ago Hello, I also have this error while using the mmlu_generative task to benchmark the llama3 model.
Command:
python3 main.py \
--model hf \
--model_args pretrained=model-path\
--tasks mmlu_humanities_generative \
--limit 3 \
--output_path output/ \
--write_outResult:
| Tasks |Version|Filter|n-shot| Metric | |Value| |Stderr|
|----------------------------|------:|------|-----:|-----------|---|----:|---|-----:|
|formal_logic | 2|none | 0|exact_match|↑ | 0|± | 0|
|high_school_european_history| 2|none | 0|exact_match|↑ | 0|± | 0|
|high_school_us_history | 2|none | 0|exact_match|↑ | 0|± | 0|
|high_school_world_history | 2|none | 0|exact_match|↑ | 0|± | 0|
|international_law | 2|none | 0|exact_match|↑ | 0|± | 0|
|jurisprudence | 2|none | 0|exact_match|↑ | 0|± | 0|
|logical_fallacies | 2|none | 0|exact_match|↑ | 0|± | 0|
|moral_disputes | 2|none | 0|exact_match|↑ | 0|± | 0|
|moral_scenarios | 2|none | 0|exact_match|↑ | 0|± | 0|
|philosophy | 2|none | 0|exact_match|↑ | 0|± | 0|
|prehistory | 2|none | 0|exact_match|↑ | 0|± | 0|
|professional_law | 2|none | 0|exact_match|↑ | 0|± | 0|
|world_religions | 2|none | 0|exact_match|↑ | 0|± | 0|
I also try to dump some intermediate result after add some log info:
a) the prompt input text: add print log for the generate_until API in lm_eval/models/huggingface.py
The following are multiple choice questions (with answers) about world religions.
Which of the following plays the most significant role in forming a child's political views?
A. The geographical area in which the child grows up
B. The child's family
C. The media to which the child is exposed
D. The child's religion
Answer:b) LLM response from self._model_generate:
The child's religionIt seems the response result looks normal, but the value of exact_match from the final result table is always 0.
Could you plase help to take a look? Thanks
AishaAlaagib
commented
3 weeks ago Hello
I have been able to solve this. I had only change the exact match to this:
def exact_match(gold, pred=None):
if pred is None and isinstance(gold, list): if len(gold) != 2: raise ValueError("If passing a single list argument, it must contain exactly two elements.") gold, pred = gold gold = str(gold).strip().upper() pred = str(pred).strip()
if not pred: print("Warning: pred is empty") return 0.0 pred_first_char = pred[0].upper() value = 1.0 if gold == pred_first_char else 0.0 return value
and I used the exact match her
dataset_path: hails/mmlu_no_train # a copy of cais/mmlu with no
auxiliary_train split
test_split: test
fewshot_split: dev
fewshot_config:
sampler: first_n
output_type: generate_until
doc_to_text: "{{question.strip()}}\nA. {{choices[0]}}\nB. {{choices[1]}}\nC.
{{choices[2]}}\nD. {{choices[3]}}\nAnswer:"
doc_to_target: "{{['A', 'B', 'C', 'D'][answer]}}"
generation_kwargs:
until:
Let me know if you have any other questions.
Best Aisha.
On Thu, 31 Oct 2024 at 11:24, yshi @.***> wrote:
Hello, I also have this error while using the mmlu_generative task to benchmark the llama3 model.
Command:
python3 main.py \ --model hf \ --model_args pretrained=model-path\ --tasks mmlu_humanities_generative \ --limit 3 \ --output_path output/ \ --write_out
Result:
Tasks Version Filter n-shot Metric Value Stderr formal_logic 2 none 0 exact_match ↑ 0 ± 0 high_school_european_history 2 none 0 exact_match ↑ 0 ± 0 high_school_us_history 2 none 0 exact_match ↑ 0 ± 0 high_school_world_history 2 none 0 exact_match ↑ 0 ± 0 international_law 2 none 0 exact_match ↑ 0 ± 0 jurisprudence 2 none 0 exact_match ↑ 0 ± 0 logical_fallacies 2 none 0 exact_match ↑ 0 ± 0 moral_disputes 2 none 0 exact_match ↑ 0 ± 0 moral_scenarios 2 none 0 exact_match ↑ 0 ± 0 philosophy 2 none 0 exact_match ↑ 0 ± 0 prehistory 2 none 0 exact_match ↑ 0 ± 0 professional_law 2 none 0 exact_match ↑ 0 ± 0 world_religions 2 none 0 exact_match ↑ 0 ± 0 I also try to dump some intermediate result after add some log info:
a) the prompt input text: add print log for the generate_until API in lm_eval/models/huggingface.py
The following are multiple choice questions (with answers) about world religions.
Which of the following plays the most significant role in forming a child's political views? A. The geographical area in which the child grows up B. The child's family C. The media to which the child is exposed D. The child's religion Answer:
b) LLM response from self._model_generate:
The child's religion
It seems the response result looks normal, but the value of exact_match from the final result table is always 0.
Could you plase help to take a look? Thanks
— Reply to this email directly, view it on GitHub https://github.com/EleutherAI/lm-evaluation-harness/issues/2279#issuecomment-2449299632, or unsubscribe https://github.com/notifications/unsubscribe-auth/AK3TI7DCFNY5IFSLKV56N7TZ6HSM7AVCNFSM6AAAAABNV5XDY2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDINBZGI4TSNRTGI . You are receiving this because you commented.Message ID: @.***>
--
DISCLAIMER: The contents of this email and any attachments are confidential. They are intended for the named recipient(s) only. If you have received this email by mistake, please notify the sender immediately and you are herewith notified that the contents are legally privileged and that you do not have permission to disclose the contents to anyone, make copies thereof, retain or distribute or act upon it by any means, electronically, digitally or in print. The views expressed in this communication may be of a personal nature and not be representative of AIMS-NEI and/or any of its Centres or Initiatives.
RawthiL
commented
3 weeks ago It is a bug in the extraction filtering. Take a look at the this log:
{"doc_id": 9, "doc": {"question": "According to Kant, morality requires us to:", "subject": "philosophy", "choices": ["perform the action that leads to the greatest total happiness.", "act only on maxims that we can will to become universal laws.", "behave only in such a way as a perfectly virtuous person would behave.", "place the interests of others above the interests of ourselves."], "answer": 1}, "target": "B", "arguments": {"gen_args_0": {"arg_0": "The following are multiple choice questions (with answers) about philosophy.\n\nPsychological egoism is:\nA. an ethical theory about how we ought to behave.\nB. a generalization concerning the way people tend to behave.\nC. a claim about human nature and the ways people are capable of behaving.\nD. none of the above.\nAnswer: C\n\nAccording to Moore’s “ideal utilitarianism,” the right action is the one that brings about the greatest amount of:\nA. pleasure.\nB. happiness.\nC. good.\nD. virtue.\nAnswer: C\n\nAccording to d'Holbach, people always act according to _____.\nA. free choices\nB. dictates of the soul\nC. necessary natural laws\nD. undetermined will\nAnswer: C\n\nAccording to Kant, morality requires us to:\nA. perform the action that leads to the greatest total happiness.\nB. act only on maxims that we can will to become universal laws.\nC. behave only in such a way as a perfectly virtuous person would behave.\nD. place the interests of others above the interests of ourselves.\nAnswer:", "arg_1": {"until": ["</s>", "\n"]}}}, "resps": [[" B"]], "filtered_resps": [" B"], "doc_hash": "c5177394044574b9c8f03867fc2e5db56e8e8904af717f33f6701af2f62c4b17", "prompt_hash": "18cd89493222e9a9fe80fd0b2beaf39dffc9abe61ff3abeb1ad50d9d33ac731c", "target_hash": "df7e70e5021544f4834bbee64a9e3789febc4be81470df629cad6ddb03320a5c", "exact_match": 0.0}it returns "exact_match": 0.0 because "filtered_resps": [" B"], is not equal to "target": "B",, note the initial space in the filtered answer, this is a normal issue, and I also observed it in BBH.
If we modify the task and templates like this:
We will get the expected result:
{"doc_id": 9, "doc": {"question": "According to Kant, morality requires us to:", "subject": "philosophy", "choices": ["perform the action that leads to the greatest total happiness.", "act only on maxims that we can will to become universal laws.", "behave only in such a way as a perfectly virtuous person would behave.", "place the interests of others above the interests of ourselves."], "answer": 1}, "target": "B", "arguments": {"gen_args_0": {"arg_0": "The following are multiple choice questions (with answers) about philosophy.\n\nPsychological egoism is:\nA. an ethical theory about how we ought to behave.\nB. a generalization concerning the way people tend to behave.\nC. a claim about human nature and the ways people are capable of behaving.\nD. none of the above.\nAnswer: C\n\nAccording to Moore’s “ideal utilitarianism,” the right action is the one that brings about the greatest amount of:\nA. pleasure.\nB. happiness.\nC. good.\nD. virtue.\nAnswer: C\n\nAccording to d'Holbach, people always act according to _____.\nA. free choices\nB. dictates of the soul\nC. necessary natural laws\nD. undetermined will\nAnswer: C\n\nAccording to Kant, morality requires us to:\nA. perform the action that leads to the greatest total happiness.\nB. act only on maxims that we can will to become universal laws.\nC. behave only in such a way as a perfectly virtuous person would behave.\nD. place the interests of others above the interests of ourselves.\nAnswer:", "arg_1": {"until": ["</s>", "\n"]}}}, "resps": [[" B"]], "filtered_resps": ["B"], "doc_hash": "c5177394044574b9c8f03867fc2e5db56e8e8904af717f33f6701af2f62c4b17", "prompt_hash": "18cd89493222e9a9fe80fd0b2beaf39dffc9abe61ff3abeb1ad50d9d33ac731c", "target_hash": "df7e70e5021544f4834bbee64a9e3789febc4be81470df629cad6ddb03320a5c", "exact_match": 1.0}see "exact_match": 1.0 at the end of the line.
I tested this on Qwen2.5-32B-Instruct-AWQ (only 50 samples) The accuracy changed from all zeros to:
| Groups |Version| Filter |n-shot| Metric | |Value | |Stderr|
|------------------|-------|------------|------|-----------|---|-----:|---|-----:|
|mmlu (generative) | 2|get_response| |exact_match|↑ |0.8351|± |0.0067|
| - humanities | N/A|get_response| |exact_match|↑ |0.8523|± |0.0136|
| - other | N/A|get_response| |exact_match|↑ |0.8231|± |0.0144|
| - social sciences| N/A|get_response| |exact_match|↑ |0.8700|± |0.0132|
| - stem | N/A|get_response| |exact_match|↑ |0.8095|± |0.0122|This is the same problem I observed in BBH, I'm planning on creaiting a PR later
Edit: Added 'take_first' to filter, it changes nothing here (in terms of results), but it breaks exact match if multiple words are going to be matched.
1436033631
commented
3 weeks ago Hi RawthiL Thanks for pointing out the missing config for the YAML file. But there are some differences in the output sequence of our model, and here is "State-sponsored prayer during school hours is prohibited, but voluntary prayer by student groups before school is allowed" after applying the above patch for the filter config.
we can see the output matches the context of D, but the exact_match is equal to 0 since the response after the filter is not equal to "D". Do you have any experience with this special response for the filter?
Thanks
{"doc_id": 0, "doc": {"question": "Which of the following best describes the balance the Supreme Court has struck between the establishment clause and the free-exercise clause?", "subject": "high_school_government_and_politics", "choices": ["Freedom of speech is protected except in certain situations, such as yelling \"fire\" in a crowded theater.", "Once a church has been recognized by the federal government, its tax-exempt status can never be revoked.", "Once Congress has created an administrative agency, that agency can be dissolved only by a constitutional amendment.", "State-sponsored prayer during school hours is prohibited, but voluntary prayer by student groups before school is allowed."], "answer": 3}, "target": "D", "arguments": {"gen_args_0": {"arg_0": "The following are multiple choice questions (with answers) about human aging.\n\nWhich of the following best describes the balance the Supreme Court has struck between the establishment clause and the free-exercise clause?\nA. Freedom of speech is protected except in certain situations, such as yelling \"fire\" in a crowded theater.\nB. Once a church has been recognized by the federal government, its tax-exempt status can never be revoked.\nC. Once Congress has created an administrative agency, that agency can be dissolved only by a constitutional amendment.\nD. State-sponsored prayer during school hours is prohibited, but voluntary prayer by student groups before school is allowed.\nAnswer:", "arg_1": {"until": ["</s>", "\n"]}}}, "resps": [[" State-sponsored prayer during school hours is prohibited, but voluntary prayer by student groups before school is allowed."]], "filtered_resps": ["State-sponsored prayer during school hours is prohibited, but voluntary prayer by student groups before school is allowed."], "doc_hash": "8f63cebd5269df80a7f6386afb6ea7266a908ffe6b72f431cf962d8dc3948358", "prompt_hash": "f63bb19b3a6c11a40c8939643328509dfd97d1b172f25a68894559a9689ba51d", "target_hash": "3f39d5c348e5b79d06e842c114e6cc571583bbf44e4b0ebfda1a01ec05745d43", "exact_match": 0.0}{"doc_id": 0, "doc": {"question": "Which of the following best describes the balance the Supreme Court has struck between the establishment clause and the free-exercise clause?", "subject": "high_school_government_and_politics", "choices": ["Freedom of speech is protected except in certain situations, such as yelling \"fire\" in a crowded theater.", "Once a church has been recognized by the federal government, its tax-exempt status can never be revoked.", "Once Congress has created an administrative agency, that agency can be dissolved only by a constitutional amendment.", "State-sponsored prayer during school hours is prohibited, but voluntary prayer by student groups before school is allowed."], "answer": 3}, "target": "D", "arguments": {"gen_args_0": {"arg_0": "The following are multiple choice questions (with answers) about human aging.\n\nWhich of the following best describes the balance the Supreme Court has struck between the establishment clause and the free-exercise clause?\nA. Freedom of speech is protected except in certain situations, such as yelling \"fire\" in a crowded theater.\nB. Once a church has been recognized by the federal government, its tax-exempt status can never be revoked.\nC. Once Congress has created an administrative agency, that agency can be dissolved only by a constitutional amendment.\nD. State-sponsored prayer during school hours is prohibited, but voluntary prayer by student groups before school is allowed.\nAnswer:", "arg_1": {"until": ["", "\n"]}}}, "resps": [[" State-sponsored prayer during school hours is prohibited, but voluntary prayer by student groups before school is allowed."]], "filtered_resps": ["State-sponsored prayer during school hours is prohibited, but voluntary prayer by student groups before school is allowed."], "doc_hash": "8f63cebd5269df80a7f6386afb6ea7266a908ffe6b72f431cf962d8dc3948358", "prompt_hash": "f63bb19b3a6c11a40c8939643328509dfd97d1b172f25a68894559a9689ba51d", "target_hash": "3f39d5c348e5b79d06e842c114e6cc571583bbf44e4b0ebfda1a01ec05745d43", "exact_match": 0.0}
It looks like you are doing zero-shot (presenting no examples prior asking the question), this results in the model not being conditioned to respond with a letter (instead an explicit response) and hence the exact match fails.
There is no way to solve that with an exact-match, you will need to create a new test definition for zero shot and probable code a different metric (like a quasi-exact-match).
If there is no important reason for you to use zero-shot, I would suggest you to add --num_fewshots 3.
1436033631
commented
3 weeks ago {"doc_id": 0, "doc": {"question": "Which of the following best describes the balance the Supreme Court has struck between the establishment clause and the free-exercise clause?", "subject": "high_school_government_and_politics", "choices": ["Freedom of speech is protected except in certain situations, such as yelling "fire" in a crowded theater.", "Once a church has been recognized by the federal government, its tax-exempt status can never be revoked.", "Once Congress has created an administrative agency, that agency can be dissolved only by a constitutional amendment.", "State-sponsored prayer during school hours is prohibited, but voluntary prayer by student groups before school is allowed."], "answer": 3}, "target": "D", "arguments": {"gen_args_0": {"arg_0": "The following are multiple choice questions (with answers) about human aging.\n\nWhich of the following best describes the balance the Supreme Court has struck between the establishment clause and the free-exercise clause?\nA. Freedom of speech is protected except in certain situations, such as yelling "fire" in a crowded theater.\nB. Once a church has been recognized by the federal government, its tax-exempt status can never be revoked.\nC. Once Congress has created an administrative agency, that agency can be dissolved only by a constitutional amendment.\nD. State-sponsored prayer during school hours is prohibited, but voluntary prayer by student groups before school is allowed.\nAnswer:", "arg_1": {"until": ["", "\n"]}}}, "resps": [[" State-sponsored prayer during school hours is prohibited, but voluntary prayer by student groups before school is allowed."]], "filtered_resps": ["State-sponsored prayer during school hours is prohibited, but voluntary prayer by student groups before school is allowed."], "doc_hash": "8f63cebd5269df80a7f6386afb6ea7266a908ffe6b72f431cf962d8dc3948358", "prompt_hash": "f63bb19b3a6c11a40c8939643328509dfd97d1b172f25a68894559a9689ba51d", "target_hash": "3f39d5c348e5b79d06e842c114e6cc571583bbf44e4b0ebfda1a01ec05745d43", "exact_match": 0.0}
It looks like you are doing zero-shot (presenting no examples prior asking the question), this results in the model not being conditioned to respond with a letter (instead an explicit response) and hence the exact match fails. There is no way to solve that with an
exact-match, you will need to create a new test definition for zero shot and probable code a different metric (like aquasi-exact-match). If there is no important reason for you to use zero-shot, I would suggest you to add--num_fewshots 3.
Got it, many thanks for your help.
RawthiL
commented
4 days ago Opened a PR to fix this : https://github.com/EleutherAI/lm-evaluation-harness/pull/2503
Hi thanks for providing such wonderful evaluation toolkit.
I was wondering why evaluation on
mmlu_generativereturns 0 accuracy whenever what models I try (pythia, qwen).I understand it as a generative version of mmlu, it can be used to evaluate base/instruct model and match the model's output to a formatted target answer ""{{['(A)', '(B)', '(C)', '(D)'][answer]}}""
My command:
Results: