---
license: cc-by-sa-4.0
language:
- en
tags:
- mixture-of-experts
- moe
- interpretability
- expert-routing
- olmoe
- finance
- model-pruning
pretty_name: OLMoE Finance vs. General Expert Routing & Specialization
size_categories:
- 1M<n<10M
configs:
- config_name: routing_traces
  data_files: routing_traces/routing_v1.parquet
- config_name: slot_analysis
  data_files: routing_traces/slot_analysis_v1.parquet
- config_name: ablation
  data_files: routing_traces/ablation_v1.parquet
- config_name: pruning
  data_files: routing_traces/pruning_v1.parquet
- config_name: coactivation
  data_files: routing_traces/coactivation_v1.parquet
- config_name: corpus_finance
  data_files: corpus/finance_matched_v1.parquet
- config_name: corpus_contrast
  data_files: corpus/contrast_matched_v1.parquet
---

# OLMoE Finance vs. General — Expert Routing & Specialization

Every routing decision made by **[OLMoE-1B-7B](https://huggingface.co/allenai/OLMoE-1B-7B-0924)** (AI2's open-weights mixture-of-experts model) across ~2.1 million token×layer events of contrasting text — **public-domain SEC filings (finance)** vs. **Wikipedia (general)** — plus the per-slot specialization analysis, ablation results, and pruning sweep derived from it.

This dataset accompanies a write-up on identifying and pruning domain-specialized experts in MoE models. **Headline finding:** ~16% of OLMoE's expert slots show strong, statistically robust finance-vs-general specialization, and masking the experts the model doesn't use for finance preserves finance-domain quality while degrading general-text quality — a clean, monotonic separation.

- 📝 **Write-up:** see the [project repo](https://github.com/miguelbetances/Mixture-of-Experts-Profiling) (`BLOG.md`)
- 💻 **Code:** https://github.com/miguelbetances/Mixture-of-Experts-Profiling

## How it was built

Two corpora were length-matched (50-token buckets) so they differ only in **domain**, not length or register-of-origin: 652 finance paragraphs (PleIAs/SEC, public-domain SEC filings) and 652 general paragraphs (Wikipedia 20231101.en). Each paragraph (~100 tokens) was run through OLMoE-1B-7B in inference mode with forward hooks on each of the 16 layers' routers. At every token the hook recorded the top-8 selected experts and their gate weights, yielding one record per (token, layer). Routing was **not** captured on prompts/questions, LLM-generated text, or copyrighted material — only declarative, openly-licensed human-written text, to isolate *domain* routing rather than *instruction* routing.

## Configs / files

Load any config by name:

```python
from datasets import load_dataset
ds = load_dataset("miguelbetances/moe-finance-specialization", "routing_traces")
```

### `routing_traces` — `routing_traces/routing_v1.parquet` (~2.14M rows)
The raw routing records — one per (token, layer).

| column | type | description |
| --- | --- | --- |
| `token_position` | int | position of the token within its paragraph |
| `token_id` | int | OLMoE tokenizer id |
| `token_str` | str | decoded token string |
| `layer` | int | layer index, 0–15 |
| `expert_indices` | list[int] | the top-8 expert ids selected at this (token, layer) |
| `expert_weights` | list[float] | gate weights for the selected experts (same order) |
| `corpus` | str | `finance` or `contrast` (= Wikipedia/general) |
| `paragraph_idx` | int | index of the source paragraph within its corpus |

### `slot_analysis` — `routing_traces/slot_analysis_v1.parquet` (1,024 rows)
Per expert-slot (16 layers × 64 experts) specialization statistics.

| column | type | description |
| --- | --- | --- |
| `layer`, `expert` | int | slot identity |
| `contrast`, `finance` | float | firing counts in each corpus |
| `contrast_rate`, `finance_rate` | float | firing rate per token |
| `log_odds` | float | smoothed log-odds (finance / contrast); + = finance-leaning |
| `total` | float | total firings |
| `z`, `p_raw`, `p_adj` | float | z-test statistic, raw and Benjamini-Hochberg-adjusted p |
| `significant` | bool | passes FDR correction |

### `ablation` — `routing_traces/ablation_v1.parquet` (25 rows)
Single-expert ablation sweep: zero out one expert's gate weight and measure the change in finance log-likelihood on a held-out probe set.

| column | type | description |
| --- | --- | --- |
| `layer`, `expert` | int | ablated slot |
| `group` | str | `finance_specialist` / `contrast_specialist` / `mid_finance` |
| `log_odds` | float | the slot's specialization score |
| `baseline_ll`, `ablated_ll` | float | mean per-token LL before/after ablation |
| `delta_ll` | float | `ablated_ll − baseline_ll` (positive = removal helped) |

### `pruning` — `routing_traces/pruning_v1.parquet` (6 rows)
Multi-expert pruning sweep: mask the top-N most contrast-leaning slots and evaluate both domains.

| column | type | description |
| --- | --- | --- |
| `n_pruned` | int | number of expert slots masked (0, 25, 50, 100, 150, 200) |
| `finance_ll`, `general_ll` | float | mean per-token LL on each held-out probe set |
| `delta_finance`, `delta_general` | float | change from baseline |
| `gap` | float | `general_ll − finance_ll` margin |

### `coactivation` — `routing_traces/coactivation_v1.parquet` (~28k rows)
Expert co-activation (PMI) per layer: how often expert pairs fire together within a token.

| column | type | description |
| --- | --- | --- |
| `layer` | int | layer index |
| `e1`, `e2` | int | the expert pair |
| `fin_count`, `con_count` | int | co-firing counts per corpus |
| `pmi_fin`, `pmi_con` | float | pointwise mutual information per corpus |
| `pmi_diff` | float | `pmi_fin − pmi_con` |

### `corpus_finance` — `corpus/finance_matched_v1.parquet` (652 rows)
Length-matched finance paragraphs.

| column | type | description |
| --- | --- | --- |
| `text` | str | the paragraph |
| `source` | str | `sec` |
| `filing_id` | str | SEC filing identifier |
| `company_cik` | str | SEC company CIK |
| `year` | int | filing year (skews toward 2012 — see limitations) |

### `corpus_contrast` — `corpus/contrast_matched_v1.parquet` (652 rows)
Length-matched general (Wikipedia) paragraphs.

| column | type | description |
| --- | --- | --- |
| `text` | str | the paragraph |
| `source` | str | `wikipedia` |
| `article_title` | str | source article title |
| `article_id` | str | Wikipedia article id |

## Limitations

- **One model, one domain pair.** OLMoE-1B-7B on SEC vs. Wikipedia. The methodology generalizes; whether the quantitative results do is an open question.
- **Corpus skew.** The finance corpus skews toward 2012 SEC filings (a sampling artifact); the contrast corpus draws from newer, lower-traffic Wikipedia articles. Each reflects one register within its domain.
- **Probes held out from analysis, not pretraining.** OLMoE saw similar text during pretraining, so the probe evaluation is a routing-stability test, not a generalization test.
- **"Pruning" = masking.** Ablation/pruning zero out gate weights; weights still reside in memory. This demonstrates the principle, not real compression.

## Licensing

This dataset bundles material under different terms; the repo is released under **CC BY-SA 4.0** to satisfy the most restrictive component (Wikipedia).

- **Finance corpus** — derived from public-domain U.S. SEC filings (via PleIAs/SEC). Public domain.
- **Contrast corpus** — derived from Wikipedia (20231101.en). CC BY-SA 4.0; attribution to Wikipedia contributors.
- **Routing traces & analysis** — produced by running OLMoE-1B-7B (Apache-2.0) over the above corpora.

## Citation

```bibtex
@misc{betances2026moefinance,
  author       = {Miguel Betances},
  title        = {OLMoE Finance vs. General: Expert Routing \& Specialization},
  year         = {2026},
  publisher    = {HuggingFace},
  howpublished = {\url{https://huggingface.co/datasets/miguelbetances/moe-finance-specialization}}
}
```