Training data: 7,170 records: the third format's training material re-answered in the
fourth format, plus 1,600 sentences that carry a scope or lifetime clause, in four languages,
from templates filled with the topology. Scoped sentences of the held-out paraphrase pool were
kept out of training.
Files:adapter_model.safetensors and adapter_config.json at the root are the released
adapter; adapter-epoch1/ and checkpoint/ are training artefacts.
Developed by: Wadood Ranjha. License: Apache-2.0.
Results (from the project's results files)
On 800 blind policy sentences (200 scope and lifetime phrases written by a separate AI
instance, attached to the third round's blind sentences): right meaning, scope and lifetime
included, 93.5 % (en 95.0, de 92.0, fr 95.0, it 92.0); by kind, a bit rate 90–97.5, requests
from a node 72.5–90 (the weakest), to a node 92.5–100, a number of hours 90–100, until further
notice 92.5–100.
On the third format's test sets it is at least as good as reader 1: 96.7 / 93.2 / 96.0 /
94.6 % (en / de / fr / it); of instruction types shown to no component it silently applies
16.7 % (reader 1: 46.3 %).
Closed loop on a stream of standing policies (three at the start, one every two simulated
hours, at most four active; 240 and 390 Erlang, three seeds): the policy table built from its
readings breaks 1.1 / 1.5 % of requests against the true table, where ignoring scope breaks
12.7 / 30.0 %, ignoring lifetime 11.9 / 6.4 %, first fit 42.0 / 38.1 %, and a reader of the
third format (no scope, no lifetime) 21.4 / 36.9 %.
In 4-bit GGUF form on a laptop GPU it reads 90.6 % of the policy sentences and its table stays
within 2.4 points of the truth.
On the third round's mixed stream of per-request instructions (390 Erlang, three seeds) it can
replace reader 1: alone it breaks 10.7 % of requests (reader 1: 10.6 %), in front of the decider
9.1 % (9.5 %), passing 16.1 % of requests to the decider instead of 11.5 %.
Template lines (by the author, partly with ChatGPT 6.1 Sol), at 16 bits: right on 34 of the owner set's 40 and on 36 of template set 1's
40 (English, German, Italian, Swedish; 15 standing policies), Swedish 9 of 10, although no
training sentence is Swedish. In 4 bits on the laptop: 31, 35, and 33 and 34 of two more sets of
40 in English and German (English 46 of 50, German 21 of 30), 61 of 70 and 17 of 20 of sets 4
and 5 (English, Swiss German, French, Italian): 211 of all 250.
The same 60 items in nine languages and scripts (written and translated by separate AI
instances; 4 bits on the laptop; ): right on 55 in English, 52–53
in German, French and Italian, 43 in Urdu script, 45 in Hindi and in Arabic script, 32 in
Urdu/Hindi in Latin letters, 14 in Arabic in Latin letters.
How to use
As reader 1 (blackfish021/apertus-lightpath-translator-v3, see its card), with this
repository's name and the fourth format's prompt (src/translate.py, SYSTEM4). A GGUF of the
adapter for llama.cpp is in blackfish021/apertus-lightpath-translator-v3-gguf
(lightpath-translator-v4-lora-f16.gguf).
Intended use and limits
Research and prototyping on a simulator; synthetic sentences; one training run; no second reader
of this format yet, so scope and lifetime are read once. Other models' role: Claude (Anthropic)
wrote the project's code and documentation under the author's direction, the sentence templates
of the training data and, as separate instances that never saw the repository, the test
sentences; no other model is part of the system. The author wrote the template lines (a test set) partly with ChatGPT 6.1 Sol (OpenAI).
Known error: "to Prague, avoid Berlin" is read as a scope (to:Prague) where an operator means
the current request's destination; the language has no word yet for a request's own end points.