blackfish021/apertus-lightpath-translator-v3

Apertus 8B lightpath translator, third format (reader 1)

A LoRA adapter that turns an operator's routing instruction for an optical transport network, in English, German, French or Italian, into one line of a small policy language that a rule engine enforces:

avoid=Zurich-Milan,Basel; via=Lyon; hops=4; km=any; path=any; prefer=load,hops; rest=none

The slots: links and nodes to avoid, a node to pass through, limits on hops and length, a candidate by its number, preferences in the order stated, and rest=more for a part the slots cannot carry. Three answers refuse: unsupported (about the route, but outside the language), no_instruction (not a routing instruction), and a name that is not a node leaves the line unresolved. Built for Hack Apertus 2026 (online, Track 2B) by the project Apertus Lightpath; this is reader 1 of its two readers.

Model details

• Base model: andreasmartin/apertus-v1.5-8b-text, the text-only conversion of swiss-ai/Apertus-v1.5-8B (accepted by the organisers for the hackathon).
• Method: QLoRA, 4-bit NF4 base, rank 32, alpha 32, 79.7 M trainable parameters, completion-only loss; one A100 80 GB, 34 minutes. Decoding is greedy and constrained by a grammar built from the network's node list, so the answer is always a line of the language.
• Training data: 5,570 instructions in four languages from templates filled with the topology (nobel-eu, 28 nodes), with the line each means; eight clause pairs trained, seven pairs and five triples held out, four instruction types absent from all training. The split was fixed and committed before any test sentence was opened.
• Files: adapter_model.safetensors and adapter_config.json at the root are the released adapter; adapter-epoch1/ and checkpoint/ are training artefacts kept for the record.
• Developed by: Wadood Ranjha. License: Apache-2.0.

Results (from the project's results files)

• Right meaning on 2,078 blind test sentences written by separate AI instances: 94.8 % (one clause 94.6, clause pairs seen in training 95.6, pairs never seen together 96.5, triples 91.2); by language 96.7 (en), 91.4 (de), 96.0 (fr), 94.8 (it) %.
• Decisions of the rule engine under its lines comply with the instruction on 99.6 % of 672 probe requests (the fine-tuned decider alone: 91.1 %; first fit: 66.8 %).
• Apertus 70B with 15 examples and no training reads 89.2 % of the same sentences.
• Weak point: of four instruction types no component was ever shown, it silently reads 46.6 % as something in the language; a second reader of the same recipe (...-translator-v3-b) catches part of that by disagreeing.
• In 4-bit GGUF form on a laptop GPU (RTX 4060, llama.cpp) it reads 93.1 % of the same sentences at 0.89 s each.

How to use

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "andreasmartin/apertus-v1.5-8b-text"
tok = AutoTokenizer.from_pretrained(base)
model = PeftModel.from_pretrained(AutoModelForCausalLM.from_pretrained(base), "blackfish021/apertus-lightpath-translator-v3")

The prompt (system text, the node list, the instruction) and the answer grammar are in the project repository, src/translate.py and src/grammar.py; without the grammar the model still answers in the format, with it the answer is guaranteed to parse. A GGUF of the adapter for llama.cpp is in blackfish021/apertus-lightpath-translator-v3-gguf.

Intended use and limits

For research and prototyping of intent-based control of optical networks on a simulator. The training and test sentences are synthetic; no operator wrote them. The language carries seven slots; everything else is refused or handed on. Trained once; no second training run for error bars. 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).

Repository: github.com/theBlackfish01/ApertusNetworkGym (track_2b/). Report: technical_report.md in that repository.