Claude Code plugin: post-train LLMs (SFT, LoRA, DPO, merging) as controlled experiments with error bars and significance tests
-
Updated
Jun 30, 2026 - Python
Claude Code plugin: post-train LLMs (SFT, LoRA, DPO, merging) as controlled experiments with error bars and significance tests
Llama-3.2-3B-Instruct distilled on DeepSeek-R1 reasoning traces to think out loud. GSM8K accuracy falls from 75.0% to 65.0% while output length more than doubles — published as a measured negative result, with the per-problem breakdown behind the headline number.
A LoRA adapter that teaches Qwen2.5-1.5B-Instruct to summarise conversations. 18.5M trainable parameters — 1.20% of the base model — tuned on DialogSum, lifting ROUGE-1 from 0.2822 to 0.4185 over prompting the same weights. Loads on top of the base model with PEFT in two lines.
Large Language Model (LLM) fine-tuning repository focused on Quantized Low-Rank Adaptation (QLoRA) and memory-efficient training. Covers 4-bit quantization, NF4, Double Quantization, LoRA adapters, supervised fine-tuning, dataset preparation, and model evaluation using Python, PyTorch, Hugging Face PEFT, TRL, and BitsAndBytes.
Qwen2.5-3B-Instruct turned into a SQL writer with a 16-bit LoRA. AST-normalised exact match rises from 7.7% to 74.3% on 300 held-out queries, and every output parses as valid SQL. The same cleaning was applied to both models, so the gain is the SQL being right, not the formatting.
LLM research plugin: 3-level generation smoke (real Qwen2.5-0.5B run) + PPL sanity eval + eval-cheating guards
A practical Large Language Model (LLM) fine-tuning repository focused on Low-Rank Adaptation (LoRA) and Parameter-Efficient Fine-Tuning (PEFT). Covers LoRA architecture, adapter configuration, dataset preparation, supervised fine-tuning, model evaluation, and efficient LLM customization using Python, PyTorch, and Hugging Face Transformers.
A LoRA adapter that steers Helsinki-NLP/opus-mt-en-es toward literary prose. 589K trainable parameters — 0.75% of the model — trained on opus_books en-es. 19.99 sacreBLEU against 17.09 for the untouched base model. 2.4 MB; loads on top of the base with PEFT.
Finetune AI Models : Reusable Hugging Face PEFT Lora and QLora finetuning scripts
Comprehensive, delta-focused study suite and interactive exam simulator for the Google Cloud Professional Machine Learning Engineer (PMLE) certification (June 1, 2026 Exam Guide).
SPICE-Verified Fine-Tuning and RL for Analog Amplifier Design : built a simulator-driven op-amp optimization agent using Qwen2.5-0.5B, LoRA fine-tuning (TRL/PEFT), and ngspice. Implemented 14 design actions, multi-step recovery, and RL based expert iteration; achieved 94.3% success across 174 test episodes, using 2.7 simulations per episode.
MTS MWS CompactFactExtraction (Innopolis Global Hack 2026) — Сжатие LLM в компактный пайплайн (0.5B-1.5B + LoRA) с нейросимвольным постпроцессингом, FastAPI и Docker.
To associate your repository with the peft-lora topic, visit your repo's landing page and select "manage topics."