On the Effect of Instruction Tuning Loss on Generalization
Anwoy ChatterjeeDept. of Electrical EngineeringIndian Institute of Technology Delhi [email protected], H S V N S Kowndinya RenduchintalaMedia and Data Science ResearchAdobe Inc., India [email protected] Sumit BhatiaMedia and Data Science ResearchAdobe Inc., India sumit.bhatia@ad · 2025
Read On the Effect of Instruction Tuning Loss on Generalization by Anwoy ChatterjeeDept. of Electrical EngineeringIndian Institute of Technology Delhi [email protected], H S V N S Kowndinya RenduchintalaMedia and Data Science ResearchAdobe Inc., India [email protected] Sumit BhatiaMedia and Data Science ResearchAdobe Inc., India sumit.bhatia@ad (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-07817.txt.
Description
Anwoy Chatterjee and three co-authors tested how loss is set when language models learn to follow instructions. Their paper, from twenty twenty-five, starts with a plain question. Most training counts loss only on the reply text, not on the prompt text. The team asked if that split is in fact the best choice. They varied the weight given to prompt and reply parts.
From those tests they built Weighted Instruction Tuning, a new way to set the weights. They tried it on five models from distinct groups and at distinct sizes. They used three training sets as well. Readers learn that a small change in the loss rule can lift how well a tuned model does on new tasks.
Catalog record
| CREATOR | Anwoy ChatterjeeDept. of Electrical EngineeringIndian Institute of Technology Delhi [email protected], H S V N S Kowndinya RenduchintalaMedia and Data Science ResearchAdobe Inc., India [email protected] Sumit BhatiaMedia and Data Science ResearchAdobe Inc., India sumit.bhatia@ad |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 47 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-07817