13 Posts getaggt mit "fine-tuning"

The final week of May 2026 (May 27–30) proved to be the grand culmination of our experimental modeling phase. After weeks of hyperparameter tuning, GPU out-of-memory crashes, and dataset revisions, our 4th Fine-Tuning cycle (fine_tuning_4) crossed the finish line.

This post reveals the technical breakthrough that resolved the chronic "over-counseling" trap, quantitative validation of our golden weight Checkpoint-4400, and the end-to-end system architecture designed for real-world deployment.

On May 21st and 22nd, the Maeumjigi Capstone team embarked on massive data generation and preparation for the 4th Fine-Tuning (fine_tuning_4).

Over the two days of May 19th and 20th, the automation of evaluation for the 3rd fine-tuning (fine_tuning_3) and the enhancement of backend/frontend features for building the actual app service proceeded in parallel.

On May 13th, after fixing the hardcoding bugs discovered yesterday, we finally resumed proper inference testing for the 2nd round fine-tuned models.

The 2nd round of fine-tuning (fine_tuning_2) parallel training, which was ambitiously started to correct the LoRA hyperparameter mistake (alpha/rank setup error), finally ended on May 12th!

May 8th was the day we were building our capstone design mid-term presentation (PPT) based on the fine-tuning results. However, while happily filling in the slides, I discovered a massive, chilling mistake.

May 7th was the day we synthesized the fine-tuning results of our 5 models (Llama3.3-70B, DSR1-70B, HCX-Omni-8B, Qwen35-9B, HCX-Text-1.5B) and proceeded with full-scale visualization and local environment setup.

On May 6th, we finally escaped the long swamp of fine-tuning! After several nights of work and hyperparameter tuning, the training for mid-size models like HCX-Omni-8B and Qwen35-9B was successfully completed.

After experiencing the OOM (Out of Memory) hell during the fine-tuning of massive 70B models, on May 4th, we attempted a bold change in strategy. By drastically reducing the batch size, we barely managed to push the 70B models (Llama3.3, DeepSeek-R1) to checkpoint-1200 and 1300, but the training time was absurdly long, resulting in terrible efficiency.

On May 3, having gained enough experience with the 8B models, our team finally challenged the ultimate kings of performance: fine-tuning massive 70B (70 billion parameters) models. We attempted to train both Llama 3.3 70B and DeepSeek-R1 70B, but as expected, we hit a massive hardware resource wall.

On April 9, using the Colab notebook we ambitiously completed yesterday, we finally attempted the first training (fine-tuning) of our artificial intelligence model (Llama 3.1). However, reality was not that easy, and we had to face numerous errors. It was also a tough day as we had to simultaneously create the presentation materials for the 6th week of the Capstone Design project.

April 8th was the day we experienced the biggest crisis and a 'laugh-and-cry' happening while working on the Maeumjigi project. It was the Claude API monthly limit exceeded & bill bomb incident.

April 1, 2026, was the day our engineering team launched the automated preprocessing pipeline for the massive counseling corpora that would form the foundation of Maeumjigi. Training a high-performance open-source Large Language Model (LLM) like Meta Llama 3.1 requires rigorous data hygiene: converting unstructured multi-party dialogues into clean, token-bounded, and special-token-aligned instruction datasets.