# Fine-tuning for a real task

A pack of 8 sessions from the MLOps Community YouTube channel, in the order to watch them. 4h 46m of video.
Page: https://mlopstalks.com/packs/fine-tuning-for-a-real-task

The fine-tuned model sounds more like your examples, but the business decision is no better. Before another training run, define the task and compare it with a smaller model or a simple baseline. A churn experiment provides useful counterevidence: adapting an LLM did not beat the conventional alternatives. Next ask whether the training objective and allowed outputs match the job. Dataset curation then separates clear instructions from convincing but wrong answers, with document and SQL examples showing the preparation work that adaptation requires. Preference training addresses a different goal when several answers are acceptable. Finish with reward-model overfitting, where a rising training score can hide a worse customer result. The sequence treats fine-tuning as a measured intervention for a particular task, not a default upgrade.

## This pack is for you if

- Fine-tuning changed your model's style, but not the task outcome you need to improve.
- You have private documents or examples but no carefully reviewed instruction-and-answer dataset.
- Your adapted model scores better during training and you need to know whether the gain survives a real task comparison.

## The talks, in order

### 1. Fine-Tuning LLMs: Best Practices and When to Go Small

Mark Kim-Huang, Preemo Inc. | 53:48 | MLOps Meetup
Video: https://www.youtube.com/watch?v=1WSUfWojoe0
Summary: https://mlopstalks.com/talks/fine-tuning-llms-best-practices-and-when-to-go-small.md

Why first: Define the capability before choosing a training method. Kim-Huang separates narrow tasks that small models already handle from cases where prompting or adaptation may help. His 2023 discussion establishes the comparison you need: measured task behavior, not a preference for owning a larger model.

### 2. LLM XGBoost: Can a Fine-Tuned LLM Beat XGBoost on Tabular Data?

Sebastian Cattes, iwt | 10:51 | MLOps Community
Video: https://www.youtube.com/watch?v=mq7sgGJ8A-g
Summary: https://mlopstalks.com/talks/llm-xgboost-can-a-fine-tuned-llm-beat-xgboost-on-tabular-data.md

Why here: A concrete baseline can defeat a plausible fine-tuning story. Cattes translates churn records into text and gets a competitive result, but XGBoost and logistic regression still outperform the LLM on his dataset. Removing a key feature also hurts both approaches. Treat this limited 2023 experiment as a reason to compare, not a verdict on all tabular tasks.

### 3. Fine-Tuning is Broken

Tanmay Chopra, Neeva | 25:27 | AI in Production 2025
Video: https://www.youtube.com/watch?v=3rk7hV8i-F8
Summary: https://mlopstalks.com/talks/fine-tuning-is-broken.md

Why here: The objective may be wrong even when the examples are good. Chopra argues for a loss aligned with the business decision and a classification head restricted to valid routes. For a stable classification task, this offers a different adaptation design from teaching a generative model to spell the label correctly; it is his argument, not a claim that instruction tuning never works.

### 4. Building and Curating Datasets for RLHF and LLM Fine-tuning

Daniel Vila Suero, Argilla | 58:51 | LLMs in Production 2023
Video: https://www.youtube.com/watch?v=Ezz_5csCJqI
Summary: https://mlopstalks.com/talks/building-and-curating-datasets-for-rlhf-and-llm-fine-tuning.md

Why here: Training pairs need independent scrutiny of the question as well as the answer. Vila Suero explains how one author can silently fill in the gaps in their own unclear prompt, and recommends collecting real queries where possible. Once the target is defined, this supplies a practical defense against teaching the model a task that users never actually asked.

### 5. Fine Tuning Llamas

Kai Davenport | 13:37 | AI in Production 2024
Video: https://www.youtube.com/watch?v=7F0tnquy6t0
Summary: https://mlopstalks.com/talks/fine-tuning-llamas.md

Why here: A document collection is not yet an instruction dataset. Davenport's 2024 Helix account turns sections into varied question-and-answer pairs and reports that improving this generation step helped fine-tunes. He also leaves retrieval open as an alternative or companion. Use the example to inspect the conversion from source text to task examples before trusting an upload-and-train interface.

### 6. Deploying LLMs on Structured Data Tasks: Lessons from the Trenches

Laurel Orr, Number Station | 31:07 | LLMs in Production 2023
Video: https://www.youtube.com/watch?v=qKBfWm3Z1oY
Summary: https://mlopstalks.com/talks/deploying-llms-on-structured-data-tasks-lessons-from-the-trenches.md

Why here: SQL gives those curation checks an executable target. Orr describes verifying queries against schemas, repairing poor questions and separating raw SQL pre-training from instruction tuning. Her 2023 ablation attributes the larger gain to connecting user requests with SQL outputs. This makes the dataset work and its contribution to task behavior inspectable, rather than relying on a fluent demonstration.

### 7. From Research to Production: Fine-Tuning & Aligning LLMs

Philipp Schmid, Hugging Face | 38:03 | AI in Production 2024
Video: https://www.youtube.com/watch?v=iwhAo_EuZRE
Summary: https://mlopstalks.com/talks/from-research-to-production-fine-tuning-aligning-llms.md

Why here: Some tasks admit several correct answers whose usefulness differs. Schmid separates instruction learning from preference-based alignment and explains DPO's direct use of comparison data. That distinction matters after the SQL example: choose data and a method for the behavior you want to change, whether correctness, structure or the kind of help an answer provides.

### 8. Tricks to Fine Tuning

Prithviraj Ammanabrolu, Databricks | 54:02 | MLOps Podcast
Video: https://www.youtube.com/watch?v=UZpUZrA0urE
Summary: https://mlopstalks.com/talks/tricks-to-fine-tuning.md

Why last: Generated feedback does not remove the need to judge the result independently. Ammanabrolu describes Tao learning from a reward model, then overfitting that model's errors as optimization continues. He also explains why repeating the same prompts adds little new information. End by checking when improvement stops being useful and what genuinely new task evidence would justify another run.