ANKR-Learning: Automated Domain-Bounded LLM Fine-Tuning with Implicit Feedback Self-Improvement and Instant Guardrail Enforcement

Read full text Summarise with VIRAT
Authors: Anil Kumar Sharma
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This document is submitted as a defensive publication. No patent rights are asserted. The submission date establishes prior art for each mechanism described. We describe ANKR-Learning, an automated pipeline for producing company-specific fine-tuned language models from operational data without a machine learning team, without cloud model vendor lock-in, and without manual annotation. The pipeline comprises five integrated mechanisms: (1) an automated training data extractor that crawls a live production codebase and generates instruction-response pairs from structural code patterns, producing 26,414 training examples from 2,179 files across 121 packages; (2) a quality filter with a minimum 0.8 threshold that enforces example quality and dataset diversity before any example enters the training corpus; (3) a domain-bounded fine-tuning system using LoRA/QLoRA adapters on open-weight base models (Llama-3-8B, CodeLlama-7B, Mistral-7B, Phi-3), which trains the model to refuse out-of-domain queries as a first-class learned behavior rather than a post-hoc filter; (4) an instant guardrail system where guardrails are implemented as rule-layer enforcement independent of model weights, enabling domain, action, compliance, and hallucination guardrails to be added or changed without retraining; and (5) an implicit feedback self-improvement loop in which deployed model interactions are monitored for acceptance/rejection signals, negative signals automatically generate additional training candidates, and each training round tightens domain accuracy without human annotation. Together, these five mechanisms constitute a novel combination not found in prior art: OpenAI fine-tuning, LangChain RAG, and open-weight fine-tuning frameworks each address at most two of the five properties; none addresses all five. The complete pipeline reduces custom enterprise LLM deployment from six months with a dedicated ML team to three weeks with a standard engineering team.

Fields of Study & Subjects

Topic Modeling Software Engineering Research Domain Adaptation and Few-Shot Learning

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
25 March 2026
Language
EN
Document type
Preprint