Human feedback that moves the model
Experts compare, rank and correct model outputs. Their judgement becomes the training signal.
- Prefer response A
- Prefer response B
RLHF and fine-tuning at MENADevs
After pre-training, a model improves through human preference. We supply the experts, run the comparisons, and fine-tune LLMs for high-value domains across coding and data science.
How the feedback loop runs
- 01Sample
We draw outputs from your model on the tasks that matter to you.
- 02Compare and rank
Experts judge responses side by side against a rubric and rank them.
- 03Correct
Where every response falls short, an expert writes the answer the model should have given.
- 04Train
Rankings and corrections go back into training as supervised data and reward signal.
What you get
Feedback from people qualified to judge the answer.
Preference ranking
Side-by-side comparisons and ranked lists of model outputs.
Trajectory review
Step-by-step review of how an agent reached its result.
Expert corrections
Rewritten responses and reasoning from domain experts.
Supervised fine-tuning
We fine-tune LLMs on the resulting data for coding and data science.
Reinforced Learning and Geographies
- Who it is for
- Teams post-training an LLM or an agent
- What you get
- Preference labels, rankings, corrections and a fine-tuned model
- Domains
- Coding and data science
- Shown in our work
- Sally, AI training and fine-tuning (2024)