autmoate/needle2-r16-e8-s42

Needle 2 — Email → Calendar Event Extraction

LoRA-fine-tuned Needle 2 (45M) that turns a normal email into a single structured calendar-event candidate.

Given a mail's text it either returns one concrete new event or refuses. It never normalizes dates and never decides who to invite — a deterministic Python layer compiles the temporal phrase and reads the recipients.

E-Mail  →  extract_event(when, title, location)  →  EventCandidate
        →  editierbare Vorschau  →  Human Approval  →  (später) Kalender

What it does

• Extracts one concrete new calendar event from an email: a verbatim temporal phrase (when), an explicit title (or none) and an optional place.
• Refuses (empty call []) when there is no single settled event: tentative scheduling, cancellations, reschedules, deadlines, past-event references, questions with a date, or unrelated mail.
• Copies, never calculates. when is the literal phrase from the message, e.g. "am 12. Oktober von 14 bis 16 Uhr" — not an ISO datetime. This keeps date resolution in deterministic, auditable code.

Scope: message text is the only input. No HTML, no attachments, no OCR.

Model details

Base	Cactus-Compute/needle2 (45M), engine 2
Method	LoRA rank 16 / alpha 32, lr 1e-4, 8 epochs, batch 8, QAT (--qat-bits auto)
Export	W4A8 .cact, 13.7 MB (merged)
Package	cactus-needle==2.0.13
Languages	German (primary), English
Domain	on-device email → calendar extraction
Artifact	needle2-r16-e8-s42.cact
SHA256	ccb63964b991e5b15ac2587721b98e40bc49354fd9471d08d8f57952f721656e

Tool schema (one tool)

extract_event(when: str, title: str = "", location: str = "") -> str

• when — the complete temporal phrase copied verbatim; all date, clock-time, end-time and timezone evidence that belongs to the event. Date-only is valid (all-day). Never ISO, never calculated.
• title — the event name exactly as written, or empty if the text has none.
• location — the place or meeting medium exactly as written; empty if none.

No event → the model returns an empty function-call list ([]).

Loading

Load with the exact schema above (argument order is part of the training contract):

import needle

tools = [
    {"name": "extract_event",
     "description": "Use only when the message contains one concrete new calendar "
                    "event that is already agreed, confirmed or announced.",
     "parameters": {"type": "object", "properties": {
         "when": {"type": "string"},
         "title": {"type": "string"},
         "location": {"type": "string"}}}},
]

agent = needle.Needle(tools=tools, system="locale: de-DE",
                      weights="needle2-r16-e8-s42.cact")

mail = "Betreff: Projektreview\n\nHallo, der Projektreview findet am 12. Oktober " \
       "von 14 bis 16 Uhr in Raum 3.14 statt."
agent.complete(mail)["function_calls"]
# [{'name': 'extract_event',
#   'arguments': {'when': 'am 12. Oktober von 14 bis 16 Uhr',
#                 'title': 'Projektreview', 'location': 'Raum 3.14'}}]

agent.reset(); agent.complete("Betreff: Termin\n\nPasst dir Dienstag 14 Uhr?")["function_calls"]
# []

The mail is presented as Betreff: <subject>\n\n<body> (message mode). In selection mode the selected text is passed alone, without the subject.

Intended use

Part of a local-first, human-in-the-loop workflow: the model proposes one candidate, a person reviews/edits it, and only then is anything written. It is designed to be paired with deterministic code that:

• compiles the when span into (start, end, all_day) against the mail's received time (relative phrases like "morgen" resolve there, not against now);
• derives invitees from structured From/To/Cc headers — the model never selects recipients;
• routes timezone/fuzzy phrases to a review step instead of guessing.

Out of scope

Calendar writes, sending mail/invitations, multi-event scheduling, reschedule or cancellation handling, timezone conversion, HTML/attachment parsing, and any network access. The model only classifies and copies evidence from the given text.

Notes

• Fine-tunes built with the Cactus platform keep no trained confidence head; confidence is reported as None. Use explicit review flags instead.
• Only German and English were trained. Other languages are untested.
• Trained for short-to-medium business mails; very long newsletters or heavy HTML-to-text artifacts may reduce accuracy.
• Always keep a human approval step before persisting anything.