blackfish021/apertus-lightpath-lora-v2-e2

Apertus 8B lightpath decider

A LoRA adapter that decides a lightpath request directly: given the request (source, destination, bit rate), the candidate paths with their state (hops, length, free spectrum, modulation, the free blocks), and an operator's instruction in plain language, it answers path=1 slot=120 or reject. Built for Hack Apertus 2026 (online, Track 2B) by the project Apertus Lightpath. In the final system it is the fallback: instructions the policy language can express are translated by a reader (blackfish021/apertus-lightpath-translator-v3) and enforced by a rule engine; instructions outside the language, or carried only in part, go to this decider, whose choice a guard checks against the part the language does carry.

Model details

• Base model: andreasmartin/apertus-v1.5-8b-text (text-only conversion of swiss-ai/Apertus-v1.5-8B).
• Method: QLoRA, 4-bit NF4 base, rank 32, alpha 32, 79.7 M trainable parameters, completion-only loss; one A100 80 GB, 138 minutes (two epochs, the second pass of two). Greedy decoding constrained to the decision line.
• Training data: 12,000 decisions from Optical Networking Gym episodes (preset jocn_benchmark, nobel-eu, seeds 101–106) labelled by rule-based teachers that follow the intent: no intent, four single intents (avoid a link, fewest hops, least loaded, fewest slots) and two intent pairs, with the avoid-link cases taught as filter-then-rank. Test episodes use other seeds.
• 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)

• Closed loop at 390 Erlang, three seeds, violations of the intent in % of requests: avoid link 3.3, fewest hops 0.6, least loaded 0.7, fewest slots 0.1, unseen wording 1.0, intent pairs never seen together 12.9 (first fit: 24.5 / 10.4 / 43.2 / 0.1 / 19.3 / 33.9; the base model without fine-tuning answers invalidly 88–98 % of the time).
• Offline agreement with the teacher's admissible set: 98.9 % on new traffic, 98.7 % on unseen wording, 85.8 % on unseen pairs.
• It follows three instruction types it was never trained on (pass through a node 94.0 %, avoid a node 83.5 %, a given candidate 84.0 % compliance, where an instruction-blind choice complies on 16.5 / 32.5 / 11.5 %) and is no better than the shortest candidate on hop or length limits.
• On the third round's probes it complies on 91.1 % of requests against 99.6 % for the reader with the rule engine; composition is its weakness (one clause 96.2 %, three clauses 76.2 %), which is why the system translates first and decides only what it cannot translate.
• A decider trained on the reader's own sentences (blackfish021/apertus-lightpath-decider-v3-e2) complies on 96.4 % of the same probes, so about two thirds of that gap is training data; in closed loop it still breaks 14.1 % of in-language requests against the reader's 0.9 %.

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-lora-v2-e2")

The prompt is rendered by src/prompt.py of the project repository from the simulator's request analysis; src/policies.py parses the answer. A GGUF of the adapter for llama.cpp is in blackfish021/apertus-lightpath-translator-v3-gguf (lightpath-decider-v2-lora-f16.gguf).

Intended use and limits

Research and prototyping on a simulator: the model ranks candidate paths the simulator has already checked for feasibility, it computes no physics. Trained on one topology and one traffic model; English instructions; synthetic data. Other models' role: Claude (Anthropic) wrote the project's code and documentation under the author's direction and the wording templates; no other model is part of the system.

Repository: github.com/theBlackfish01/ApertusNetworkGym (track_2b/).