unslothai/unsloth

Evaluation loss becomes constant

Open

#1,067 opened on Sep 27, 2024

View on GitHub
 (7 comments) (0 reactions) (0 assignees)Python (5,658 forks)batch import
currently fixinggood first issue

Repository metrics

Stars
 (64,271 stars)
PR merge metrics
 (Avg merge 3d 15h) (525 merged PRs in 30d)

Description

Hi!

I am testing out unsloth to fine tune llama 3.1 8B instruct and following your notebook here.

One exception is that I have added an eval set. What is really strange is that the eval loss locks up to a specific value after around 300 steps. I mean down to the last decimal, not just flattening out. The training loss looks fine and as expected.

I have changed many parameters and tried different things but it always happens. Any idea on what can cause this?

Contributor guide

Research direction
Examine the evaluation loop in the unsloth library to understand why eval loss becomes constant after ~300 steps. Check if the evaluation dataset is being cached or if gradients are being disabled incorrectly. Look for issues with data shuffling or fixed eval batches.
Tech stack
pythonpytorch
Domain
machine learningai
Issue type
Bug
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
3
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
Half day
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
Active
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
Mostly clear
Prerequisites
PythonPyTorch
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
40