fblgit/simple-math

Simple Math: 2+2=4 -1=3 (LoLo: Learning Only Logical Operations)

Just like my teacher gave me homework, i thought maybe we can also add some of these basics on the trainings of our models.

It was created with very simple code that is in the repo, if you add more complex operations and so.. please share the code :D thank you

Current Code Version: 20240127.fblgit (A modification over @win10 for progressive and DPO operation)

Does it Works?

34BEAGLES Evaluation:

hf (pretrained=/data/models/UNA-34Beagles-v1-final,dtype=bfloat16,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: auto (8)
|    Tasks     |Version|Filter|n-shot| Metric |Value |   |Stderr|
|--------------|-------|------|-----:|--------|-----:|---|-----:|
|arc_challenge |Yaml   |none  |    25|acc     |0.7039|±  |0.0133|
|              |       |none  |    25|acc_norm|0.7321|±  |0.0129|
|truthfulqa_mc2|Yaml   |none  |     0|acc     |0.7387|±  |0.0141|

hf (pretrained=/data/models/UNA-34Beagles-v1-final,dtype=bfloat16,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: auto
|Tasks|Version|  Filter  |n-shot|  Metric   |Value |   |Stderr|
|-----|-------|----------|-----:|-----------|-----:|---|-----:|
|gsm8k|Yaml   |get-answer|     5|exact_match|0.6399|±  |0.0132|

|      Groups      |Version|Filter|n-shot|Metric|Value |   |Stderr|
|------------------|-------|------|-----:|------|-----:|---|-----:|
|mmlu              |N/A    |none  |     0|acc   |0.7477|±  |0.1079|
| - humanities     |N/A    |none  |     0|acc   |0.7188|±  |0.0855|
| - other          |N/A    |none  |     0|acc   |0.7950|±  |0.1057|
| - social_sciences|N/A    |none  |     0|acc   |0.8297|±  |0.0664|
| - stem           |N/A    |none  |     0|acc   |0.6641|±  |0.1291|

34BEAGLES-MATH Evaluation

hf (pretrained=/data/models/34BeaglesMath-v1,dtype=bfloat16,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: auto
|Tasks|Version|  Filter  |n-shot|  Metric   |Value |   |Stderr|
|-----|-------|----------|-----:|-----------|-----:|---|-----:|
|gsm8k|Yaml   |get-answer|     5|exact_match|0.6505|±  |0.0131|

hf (pretrained=/data/models/34BeaglesMath-v1,dtype=bfloat16,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: auto (8)
|    Tasks     |Version|Filter|n-shot| Metric |Value |   |Stderr|
|--------------|-------|------|-----:|--------|-----:|---|-----:|
|arc_challenge |Yaml   |none  |    25|acc     |0.7090|±  |0.0133|
|              |       |none  |    25|acc_norm|0.7329|±  |0.0129|
|truthfulqa_mc2|Yaml   |none  |     0|acc     |0.7378|±  |0.0141|

|      Groups      |Version|Filter|n-shot|Metric|Value |   |Stderr|
|------------------|-------|------|-----:|------|-----:|---|-----:|
|mmlu              |N/A    |none  |     0|acc   |0.7524|±  |0.1045|
| - humanities     |N/A    |none  |     0|acc   |0.7307|±  |0.0846|
| - other          |N/A    |none  |     0|acc   |0.7937|±  |0.1029|
| - social_sciences|N/A    |none  |     0|acc   |0.8274|±  |0.0667|
| - stem           |N/A    |none  |     0|acc   |0.6708|±  |0.1236|

But it gets better, because when increasing length and complexity, the marks are even superior:

|Tasks|Version|  Filter  |n-shot|  Metric   |Value |   |Stderr|
|-----|-------|----------|-----:|-----------|-----:|---|-----:|
|gsm8k|Yaml   |get-answer|     5|exact_match|0.6611|±  | 0.013|

On a 3.20% GSM Improvement compared to its base model.

Note to contributors:

thank you to those contributing on the experiment with beautiful commits and good spirit

• Feel free to contribute on the readme Evaluation tests.
• Lets aim to build an ablation & paper together. All contributors will be cited.

Versions

27.01.24 Added new code to generate the dataset, seed 42 and now also generates DPO.
24.01.24 Added gradual complexity on a separate script
20-23.01.24 Multiple contributions with operations and increased complexity on the main generator script.

Citations

If you use Simple Math o train your model, please cite on the modelcard or the paper.

@misc{simplemath,
  title={Simple-Math: 2+2=4 4-1=3}, 
  author={Xavier Murias},
  year={2024},
  publisher = {Juanako.AI},
  journal = {HuggingFace repository},
  howpublished = {\url{https://huggingface.co/datasets/fblgit/simple-math}},
}