flex-math-5504

flex-math-5504

A pruned and distilled variant of allenai/Flex-math-2x7B-1T with a variable-width expert MLP. Expert 1 has been pruned from the full 11,008 intermediate size down to 5504 (50% of original width), then recovered via knowledge distillation.

For full details, see the blog post.

How to Use

This repo includes a modeling_pruned_flex_olmo.py file that handles the variable-width expert architecture. Just load with trust_remote_code=True and it works like any other HuggingFace model:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("hbfreed/flex-math-5504", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("allenai/Flex-math-2x7B-1T")

input_text = "Solve: What is 15% of 200?"
inputs = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

The tokenizer is the same as the base model's.

How It Was Made

1. Structured pruning: Neuron importance scores were computed on math-specific data (GSM8k, Metamath, TuluMath subsets). The least important neurons in Expert 1's gate/up/down projections were removed, reducing intermediate size from 11,008 to 5504.
2. Knowledge distillation: The pruned model was retrained for ~228M tokens using the top-128 logprobs from the full-sized teacher model. Distillation data: hbfreed/flexolmo-math-logprobs.

Math-calibrated importance analysis was used — 58% of the top-2048 neurons differ between math-calibrated and general-calibrated rankings.

Benchmark Results

All Variants

License

Apache 2.0 (same as base model)