---
license: apache-2.0
task_categories: [text-generation]
tags: [dynamic-chunking, tokenization, hnet, boundaries, training-dynamics]
---

# H-Net chunker boundary probes

Longitudinal boundary decisions for 31 H-Net runs, logged on a **fixed, byte-identical**
FLORES+ probe at every checkpoint. This is the raw material for studying *when* a learned
segmentation stabilises.

## Layout

`<run>/{step:06d}__{lang}.npz`, plus `<run>/probe_text.jsonl` (the raw probe text, so byte
offsets can be aligned to external gold data).

- **40 log-spaced steps**: 0, 1, 2, 4, 8, 16, 32, 64, 128, 200, then every 200 to 6000.
  The early log-spacing is deliberate — uniform 200-step spacing misses the sharp early
  transitions, and it cannot be recovered after a run.
- **12 languages**: de, el, en, fi, hi, my, ru, ta, th, tr, vi, zh
- 481 files per run (`jitter_*` runs have 157; they were logged at a coarser cadence).

### npz keys

| key | shape | meaning |
|---|---|---|
| `mask` | `(L,) uint8` | 1 = a unit boundary falls before this byte |
| `score` | `(L,) float32` | the router's continuous boundary probability |
| `sent_offsets` | `(n_sent+1,) int64` | byte offset of each probe sentence |
| `bpb` | scalar | per-language bits-per-byte at this checkpoint |
| `step` | scalar | training step |
| `b_pad`, `p_pad`, `pad_mask` | `(128, 315)` | per-sentence padded views |

**Known issue:** `mask` and `mask_level1` (and `score`/`score_level1`) are byte-for-byte
identical — the two-level logging wrote the same level twice, so per-level word-vs-morph
typing cannot be done from these dumps. For Thai, `len(mask)` exceeds the probe byte length
by 21 bytes. Both are logging bugs, recorded here rather than silently patched.

## Caveats worth knowing before you use these

- Boundaries are **phase-locked to a seed-arbitrary sub-character byte position**. Chunk
  starts land on a UTF-8 character boundary 29–31% of the time against a 33% chance rate
  (verified independently against the span dumps, not just the mask).
- 20 of 36 language-seed pairs have **not** converged by step 6000, so a step-6000 analysis
  is measuring an unconverged segmentation for the low-resource half.
- Probe text decoded with `errors="ignore"` no longer byte-aligns to the mask for truncated
  sentences (up to 35/128 for `my`); drop those rather than assume alignment.

Companion [models](https://huggingface.co/AdaptiveChunking/hnet-chunking-pilots) and
[results](https://huggingface.co/datasets/AdaptiveChunking/hnet-chunking-results).