| | --- |
| | language: |
| | - en |
| | license: apache-2.0 |
| | library_name: transformers |
| | tags: |
| | - pruned_flex_olmo |
| | - custom_code |
| | - math |
| | - pruned |
| | - distilled |
| | - mixture-of-experts |
| | base_model: allenai/Flex-math-2x7B-1T |
| | pipeline_tag: text-generation |
| | --- |
| | |
| | # flex-math-2048 |
| |
|
| | A pruned and distilled variant of [allenai/Flex-math-2x7B-1T](https://huggingface.co/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 **2048** (19% of original width), then recovered via knowledge distillation. |
| |
|
| | | | | |
| | |---|---| |
| | | **Total Parameters** | 8.1B | |
| | | **Expert 1 Parameters** | 0.8B | |
| | | **Expert 1 Width** | 2048 (19%) | |
| | | **Base Model** | allenai/Flex-math-2x7B-1T (11.6B params) | |
| |
|
| | For full details, see the [blog post](https://hbfreed.com/2026/01/28/variable-flexolmo.html). |
| |
|
| | ## 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: |
| |
|
| | ```python |
| | from transformers import AutoModelForCausalLM, AutoTokenizer |
| | |
| | model = AutoModelForCausalLM.from_pretrained("hbfreed/flex-math-2048", 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 2048. |
| | 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](https://huggingface.co/datasets/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 |
| |
|
| | | Model | GSM8K | MATH | Math2 | |
| | |---|---|---|---| |
| | | No-expert baseline (7.3B) | — | — | 8.1 | |
| | | **flex-math-2048** | **44.3** | **13.9** | **29.1** | |
| | | Full teacher (11.6B) | 69.7 | 35.4 | 52.5 | |
| |
|
| | ### All Variants |
| |
|
| | | Model | Total Params | Expert Width | GSM8K | MATH | Math2 | |
| | |---|---|---|---|---|---| |
| | | [flex-math-8192](https://huggingface.co/hbfreed/flex-math-8192) | 10.5B | 8192 (74%) | 70.1 | 31.3 | 50.7 | |
| | | [flex-math-5504](https://huggingface.co/hbfreed/flex-math-5504) | 9.5B | 5504 (50%) | 66.6 | 26.8 | 46.7 | |
| | | [flex-math-2048](https://huggingface.co/hbfreed/flex-math-2048) | 8.1B | 2048 (19%) | 44.3 | 13.9 | 29.1 | |
| |
|
| | ## License |
| |
|
| | Apache 2.0 (same as base model) |
| |
|
