Department for Education funding for English lessons has dropped by equivalent of £117m despite rising demand, exclusive data reveals

Our work with post-16 & adults (ESOL) Working with partners to break down language barriers, to education, in the workplace, and across the criminal justice system. Discover our work on late arrivals ...

This blog features Slough & Langley College, a finalist in the 2025/26 Beacon Award for Excellence in ESOL, which supports adult learners of English to live and work in the UK thanks to its innovative, wraparound ESOL programme.

The barriers to accessing ESOL teaching and recommendations for an English language strategy which aims for universal fluency.

Course rep training products include our main training as well as specific materials for students in supported education and ESOL courses.

eSol Global is urgently hiring a Forward Deployed AI Engineer to support enterprise AI initiatives within the airline and aviation sector

Thanks for the detailed response, @woongworks — and great news on the eval runner. Looking forward to the GitHub commit. A few follow-up thoughts based on your replies: **On multi-seed runs:** The computational cost is real, but even 3 seeds (vs 1) cuts the variance interpretation problem substantially. One practical middle ground: run 5 seeds only on the datasets showing the largest scaffold/random gap (FreeSolv, ESOL) and keep 1 seed on the rest. That focuses compute where the uncertainty actually matters. **On the local-vs-scaffold label insight for QM9:** This generalizes in a useful way. For any synthetic benchmark, it’s worth asking: “Is this label determined by local context, or by global structure?” Local-label tasks will always show smaller split-strategy sensitivity. This could be a useful diagnostic criterion for anyone designing evaluation datasets from scratch — scaffold/scaffold-equivalent splits are most discriminating when labels require global or distributional reasoning. **On the eval runner:** If you include scaffold split as the default evaluation mode and random split as an explicit flag, that naturally nudges the community toward better practice without preventing legacy comparisons. Would suggest outputting both scores side-by-side by default so the gap is visible in every run. Will keep an eye out for the commit link. Thanks for engaging on this — exactly the kind of methodology discussion the field benefits from.

Thanks for waiting. The code is public now: GitHub - ByungwoongYoo/contamcheck: Detect provenance shortcuts (source, batch, campaign, instrument, lab) that inflate random-split cross-validation on labelled tabular datasets. · GitHub Your three suggestions (5 seeds on FreeSolv and ESOL only, a local-vs-global label check, and scaffold split as the default while still printing the random-split score) are next on my list, and I’ll post back here when they’re in. Out of curiosity, what kind of data are you planning to run it on?

Mair Godley claims Essex councillor Carl Johnson told her she was ‘facilitating the invasion’ by giving language lessons