AkanaYB/saas-erp-support-qwen3-4b-adapter-f

Capstone support: accepted candidate F

Experimental Russian SaaS / fictional ERP / telecom support LoRA adapter. Published project artifact; project-specific licensing has not been selected. The upstream Base license does not automatically license the authored adapter or data. Code and launch tutorial: https://github.com/Akana92/SaaS-ERP-CRM-Support-Client

Base and training

Base: Qwen/Qwen3-4B-Instruct-2507, revision cdbee75f17c01a7cc42f958dc650907174af0554 (download separately; no base weights here). Clean-base QLoRA NF4, rank 16, alpha 32, dropout .05, all linear layers, two epochs / 558 optimizer steps. Training set: 430 synthetic train dialogues, 1113 target turns, 13 local shards. Pending human review. Use the project's pinned runtime and PEFT loader with the adapter files here; the copied adapter configuration points to the pinned Hub base instead of a local path. Loading this adapter alone does not reproduce LangGraph, ERP context, or server decisions.

Evidence and limitations

AI-assessed development comparison: Base 10/100 vs F 49/100 fully successful dialogues, 225 turns, NF4. Status: QUALIFIED. Success requires the five labels and semantic response criteria across all turns. This is not human-verified or final-test accuracy: human review is absent, final sealed test was not run, full_success_rate=null. Live BF16 screenshots are not a new evaluation. Do not claim 90% quality or production readiness. Real ERP integration, production authorization, and measured operator time savings are not provided.

Provenance / license scope

BASE_MODEL_LICENSE is the unchanged Apache-2.0 license from the pinned Qwen base snapshot. It applies to upstream Qwen assets (including copied tokenizer assets), and is not a license grant for this project's adapter. Upstream source: https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507/tree/cdbee75f17c01a7cc42f958dc650907174af0554 manifest.json records source and export hashes. Only the COPY of adapter_config.json changes base_model_name_or_path and revision; originals stay frozen. No optimizer states, checkpoints, evaluation questions, gold labels, raw reviews, isolation receipt, or secrets are included. This is not a complete experimental reproduction bundle. See the code repository's docs/ML_ASSETS.md.