README.md

---
pretty_name: Nomos V1 Training Pairs
license: other
license_name: mixed-source-research-preview
language:
  - en
task_categories:
  - sentence-similarity
size_categories:
  - 10K<n<100K
tags:
  - text
  - tool-routing
  - agent-routing
  - retrieval
  - sentence-transformers
  - nomos
configs:
  - config_name: default
    data_files:
      - split: train
        path: train.jsonl
---

# Nomos V1 training pairs

This is the **40,181-row training-pair snapshot for the final specialist branch** of [nomos-v1-nano-g1](https://huggingface.co/yafitzdev/nomos-v1-nano-g1), a CPU-runnable embedding model that ranks an agent's legal tools for its next step. Each row contains the exact two text fields passed to the branch's sentence-transformer trainer:

| Column | Meaning |
|---|---|
| `anchor` | Serialized agent objective and decision state (the query). |
| `positive` | Serialized metadata for one acceptable tool (the candidate). |

The model on the Hub is a **90% base-router / 10% specialist weight interpolation**. This dataset snapshots the specialist branch's final training input. The base router was trained in earlier stages; **this dataset alone does not reproduce the entire released checkpoint**. It contains no evaluation split or negative candidates.

## Size and composition

| Source cohort | Training rows |
|---|---:|
| Generic portability replay | 20,000 |
| Frozen agentic states | 4,185 |
| ToolRet-derived agentic states | 4,096 |
| Agentic transitions | 3,400 |
| Agentic contrasts | 3,400 |
| Balanced hard subset | 5,100 |
| **Total** | **40,181** |

The 40,181 rows contain **35,812 distinct `(anchor, positive)` pairs** and **33,549 distinct anchors**. Repeated rows are retained because the trainer consumed the selected rows without deduplication. Cohorts are concatenated in the order above, with rows in their original source order.

The ToolRet-derived cohort is based on [ToolRet-Training-20w](https://huggingface.co/datasets/mangopy/ToolRet-Training-20w). The other cohorts are Nomos-generated routing states. The balanced hard subset was selected from a larger scaling cohort. The text contains synthetic task and tool descriptions, including examples derived from an external tool-retrieval dataset.

## Selection and provenance

The snapshot was rebuilt with the training code's `tools.train_dense_router._pairs` function and checked against the specialist's `nomos_training_manifest.json`:

1. Read the six source JSONL files in training-manifest order.
2. Keep rows with `evaluation_partition == "train"`, excluding `task_kind == "verify"`.
3. Require at least one `label.acceptable_tools` entry and resolve its first tool in the row's registry.
4. Serialize the row with `nomos.dense_router.query_document` and that tool with `nomos.dense_router.candidate_document`.
5. Keep at most 20,000 selected rows per input file.

The branch used `MultipleNegativesRankingLoss`, one epoch, batch size 64, learning rate 2e-6, and seed 20260824. The training code used a `NO_DUPLICATES` batch sampler; repeated rows in this file therefore do not imply identical examples were placed together in the same batch.

[`provenance.json`](https://huggingface.co/datasets/yafitzdev/nomos-v1/blob/main/provenance.json) records the source-file SHA-256 hashes and byte sizes, exact row ranges and counts, extraction settings, model checkpoint reference, and the SHA-256 of `train.jsonl`. It identifies the **local source snapshots**, which are not included here. The published model also includes its inherited training manifest and final interpolation manifest.

## Use

```python
from datasets import load_dataset

pairs = load_dataset("yafitzdev/nomos-v1", split="train")
print(pairs[0]["anchor"])
print(pairs[0]["positive"])
```

For sentence-transformer training, use the `anchor` and `positive` columns as the positive query/candidate pair. The dataset is intended for inspection and research on tool-routing embeddings. It should not be treated as an independent test set or as a complete record of all earlier training stages.

## Evaluation and limits

The [model card](https://huggingface.co/yafitzdev/nomos-v1-nano-g1) reports frozen routing evaluations separately. No benchmark items are packaged here. The data is predominantly synthetic and shaped by the Nomos serializer and source registries, so performance on new agents and tool inventories must be measured independently.

## Source and reuse

This is a mixed-source research snapshot. The ToolRet-derived portion cites the upstream [dataset](https://huggingface.co/datasets/mangopy/ToolRet-Training-20w) and [project](https://github.com/mangopy/tool-retrieval-benchmark). The `mixed-source-research-preview` label describes the bundle's status; it is **not a blanket permission to reuse upstream material**. Review the terms of each upstream source before redistribution or commercial use.