fdyrd/QwenMath-0.5B

QwenMath

A generation LLM which can solve math problems.

Training Statistics

training-method: lora
training-time: "5:42"
data-size: 500
epoch: 3
total_flos: "1372250GF"
train_loss: 0.6441
train_samples_per_second: 4.385
train_steps_per_second: 0.544

Validation Set Performance

Dataset used: test split of fdyrd/MATH. Metric: accuracy

Level	Algebra	Intermediate Algebra	Prealgebra	Precalculus	Number Theory	Geometry	Counting & Probability	Average
Level 1	0.541 : 135	0.192 : 52	0.477 : 86	0.228 : 57	0.467 : 30	0.263 : 38	0.359 : 39	0.361
Level 2	0.323 : 201	0.109 : 128	0.367 : 177	0.044 : 113	0.38 : 92	0.134 : 82	0.248 : 101	0.229
Level 3	0.291 : 261	0.046 : 195	0.308 : 224	0.0 : 127	0.262 : 122	0.088 : 102	0.16 : 100	0.165
Level 4	0.18 : 283	0.024 : 248	0.22 : 191	0.009 : 114	0.169 : 142	0.064 : 125	0.09 : 111	0.108
Level 5	0.088 : 307	0.004 : 280	0.104 : 193	0.0 : 135	0.136 : 154	0.023 : 132	0.065 : 123	0.06
Average	0.285	0.075	0.295	0.056	0.283	0.114	0.184	0.166

Test Set Performance

[
  {
    "dataset": "MATH500",
    "url": "https://huggingface.co/datasets/qq8933/MATH500",
    "accuracy": 0.286
  },
  {
    "dataset": "GSM8K",
    "url": "https://huggingface.co/datasets/openai/gsm8k",
    "accuracy": 0.382
  }
]