---
license: mit
language:
  - sa
task_categories:
  - token-classification
  - other
tags:
  - sanskrit
  - samasa
  - sandhi
  - compounding
  - vigraha
  - vyakarana
  - boundary-detection
---

# Sanskrit Sandhi + Samas Boundary Dataset — V3

The canonical merged training dataset for the **sanskrit-sandhi-boundary** model:
sentence-level **external sandhi** (corpus) + grammar-generated **samas**
(compounds, all 7 types) in one file.

## Composition

| Source | Rows | Description |
|--------|------|-------------|
| sandhi_corpus | 741,803 | running-text word boundaries (external sandhi) from `sanskrit-sandhi-boundaries-v2` |
| samas | 258,408 | grammar-generated compounds (7 types, laukik + alaukik vigraha) from `sanskrit-samas-v1` |
| **Total** | **~1,000,211** | one canonical training file |

## Design rationale

The splitter's task at a seam is **phonetically identical** for external sandhi
and internal compound boundaries — the model must fire at *every* valid seam and
stay quiet inside single words. Merging (not split-training) lets the shared
phonetics reinforce both, while the samas rows teach stem-co-occurrence for the
identity-concat class (no phonetic signal) and the corpus rows keep the real
distribution. See `benchmarks/PLAN_SAMAS.md`.

## Schema (per row)

| Field | Type | Description |
|-------|------|-------------|
| `seq` | string | SLP1 surface (input to the model) |
| `bnd` | list[int] | Boundary positions (last-char indices of each piece except the last) |
| `n_boundaries` | int | len(bnd) |
| `source` | string | `sandhi_corpus` or `samas` |
| `samasa_type` | string (opt) | one of 7 types, for samas rows |
| `stem1`, `stem2` | string (opt) | constituent stems, for samas rows |
| `verified` | bool | always true |

## Samasa type coverage (in the samas portion)

| Type | Meaning | n |
|------|---------|---|
| tatpuruṣa | determinative (X of Y) | 62,893 |
| karmadhāraya | appositional (X that is Y) | 62,893 |
| dvandva | copulative (X and Y) | 59,924 |
| bahuvrīhi | exocentric "having X" | 60,172 |
| avyayībhāva | adverbial | 7,097 |
| amreḍita | reduplicative | 5,091 |
| dvigu | numerical | 338 |

## Rebuilding

```bash
# samas dataset (deterministic, seed 42)
python3 benchmarks/build_samas_data.py --out data/samas_V1.jsonl

# merged dataset (deterministic, seed 42)
python3 benchmarks/merge_datasets.py \
    --sandhi data/train_V2_full.jsonl \
    --samas data/samas_V1.jsonl \
    --out data/train_V3_sandhi_samas.jsonl
```

## Related

- Samas-only: [`sanskrit-samas-v1`](https://huggingface.co/datasets/CodeIsAbstract/sanskrit-samas-v1)
- Sandhi-only: [`sanskrit-sandhi-boundaries-v2`](https://huggingface.co/datasets/CodeIsAbstract/sanskrit-sandhi-boundaries-v2)
- Benchmark: [`sanskrit-sandhi-boundary-v2`](https://huggingface.co/CodeIsAbstract/sanskrit-sandhi-boundary-v2) (model)