Hi everyone,<\/p>\n
I’ve been working on a data-centric pipeline for constructing SFT and KTO datasets for small language models, targeting models ranging from a 1.58B ternary model up to 12B parameters (with a particular focus on the 1.5B\u20134B range), using an Italian tool-calling assistant (“Liara”) as a case study.<\/p>\n
Instead of focusing on model architecture, the goal is to reduce common failure modes through dataset construction itself:<\/p>\n
- \n
- tool over-calling<\/li>\n
- style collapse<\/li>\n
- excessive verbosity<\/li>\n
- semantic redundancy<\/li>\n
- memory inconsistencies<\/li>\n<\/ul>\n
The pipeline currently includes:<\/p>\n
- \n
- typed validation outcomes (PASS \/ Soft Reject \/ Hard Reject \/ Warning)<\/li>\n
- semantic + structural deduplication<\/li>\n
- multi-teacher generation<\/li>\n
- dataset lineage and versioning<\/li>\n
- regression set<\/li>\n
- dataset health dashboard<\/li>\n
- capability-based dataset profiling for different model sizes<\/li>\n
- typed routing into SFT, KTO-negative, or discard<\/li>\n
- Soft Reject examples are not discarded by default: they undergo additional validation and, if confirmed, are reused as KTO-negative examples rather than being treated as unusable data.<\/li>\n<\/ul>\n
The current specification describes the methodology. The implementation is underway, and the experimental validation is currently running.<\/p>\n
I’d love feedback from people who have built or maintained instruction datasets:<\/p>\n
- \n
- Which parts seem genuinely useful?<\/li>\n
- Which ideas already exist in other pipelines?<\/li>\n
- What ablation studies would you expect before considering this publishable?<\/li>\n<\/ul>\n
I’m currently generating the gold seed dataset, which is the most time-consuming part of the pipeline and is expected to take around 10 days at the planned scale. Once that’s complete, I’ll publish the implementation, the ablation results, and the evaluation so the methodology can be assessed based on experimental evidence rather than design alone.<\/p>\n
In the meantime, I’d really appreciate any feedback or suggestions on the pipeline itself.<\/p>\n<\/div>\n
submitted by \/u\/Key-Outcome-2927 <\/a>
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
