ChronoNano 1.0 10M
ChronoNano 1.0 10M is a compact, purpose-built neural model for parsing natural-language temporal expressions. It does not generate text. Its job is to convert phrases such as tomorrow, last 3 months, next Friday, or between 3 and 5 PM tomorrow into a structured temporal representation that can be deterministically resolved to a DateTimeRange.
The model reads UTF-8 bytes augmented with neutral lexical features and hashed byte n-grams. A 12-layer attention-only encoder builds the input representation. A 5-layer attention-only autoregressive decoder produces a fixed 40-slot semantic structure, while a separate 14-role pointer subsystem extracts literal values such as numbers, dates and times from the original input. Beam decoding can generate multiple structured candidates for downstream selection.
Key parameters
| Parameter | Value |
|---|---|
| Trainable parameters | 10,733,483 |
Model width (d_model) |
320 |
| Attention heads | 8 |
| Encoder layers | 12 |
| Decoder layers | 5 |
| Maximum input | 256 UTF-8 bytes + CLS |
| Input byte vocabulary | 258 IDs |
| Lexical feature vocabulary | 53 |
| Hashed byte n-grams | 2–5 bytes |
| Semantic output slots | 40 |
| Structured decoder vocabulary | 245 tokens |
| Extractive pointer roles | 14 |
| Semantic contract | V2 |
Output
ChronoNano produces two complementary outputs:
- a fixed 40-token semantic structure describing the temporal meaning;
- pointer predictions (
presence,start,end) for 14 extractive roles used to recover literal values from the input.
The neural output is intentionally separated from calendar arithmetic. A deterministic runtime resolver combines the decoded semantics with a reference date/time and converts them into a half-open interval:
[from, to)