I’ve been deep-fine-tuning small local models (BitNet 1.58, Qwen 1.7B\/4B, Gemma-4) into a real on-device assistant with actual tool-calling. The hard part was never the training loop, it’s the data<\/strong>. Scraped<\/p>\n ShareGPT dumps collapse small models into formulaic slop and teach tool syntax that doesn’t match your runtime. And for full\/deep fine-tuning<\/strong> (not LoRA), bad data is fatal: a small model trained on redundant, single-voice data monocultures<\/em> hard.<\/p>\n So I built a pipeline where every example has to survive a set of hard gates<\/strong> before it’s allowed in.<\/p>\n Sharing how it works, because I rarely see people talk about the controls<\/em>, only the volume.<\/p>\n The core: hand-written “gold seed” \u2192 multi-teacher expansion<\/strong><\/p>\n – The seed is hand-written<\/strong>, one example at a time, in a neutral model-agnostic format<\/strong> ({messages, tools}).<\/p>\n – It’s rendered per dialect<\/strong>: ChatML for Qwen\/BitNet, native Gemma tool-call format for Gemma. Same data, correct syntax for each target.<\/p>\n – From a curated seed it expands to hundreds of thousands of examples on demand<\/strong> \u2014 volume comes from multiple teacher models of different families<\/strong> (anti style-collapse), each example tagged with the teacher that produced it. You scale to the size you actually need.<\/p>\n The gates (this is the value)<\/strong><\/p>\n – Anti-formula<\/strong>: blocks over-used openers\/closers at ingest + global frequency caps; any phrase repeated too often anywhere<\/em> gets flagged. This is what stops a full fine-tune from collapsing into one voice.<\/p>\n – Semantic dedup (BGE-M3)<\/strong>: near-duplicate detection, not byte-match. On one 9k combinatorial corpus it found 43% near-duplicates<\/strong>, paraphrases expansion would’ve amplified. Keeps 1 per cluster, with a coverage<\/strong> guard<\/strong> that never drops the sole example teaching a capability.<\/p>\n – Flow gate<\/strong>: multi-turn tool-call integrity: every assistant tool_call is followed by exactly<\/em> its results, no orphans, no pending calls, ends with a real answer. <\/p>\n – Dialect gate<\/strong>: every tool-call is round-tripped through each target model’s syntax and rejected if it<\/strong> doesn’t parse back identical<\/strong>. train == runtime, guaranteed.<\/p>\n – Coverage<\/strong>: every tool taught above threshold; catalog-sync rejects phantom (hallucinated) and untaught tools.<\/p>\n – Vision routing<\/strong>: vision examples go only<\/em> to vision-capable models; text-only models never see image context they can’t use at inference.<\/p>\n – Judge routing<\/strong> \u2014 teacher outputs that pass go to SFT; the ones that fail become KTO negatives<\/strong> (preference signal, not trash).<\/p>\n Built for deep fine-tuning, at scale<\/strong><\/p>\n The whole point of the gates is that you can full-fine-tune a small model<\/strong> on this without it collapsing and expand a small hand-verified seed to 100k+ examples<\/strong> while keeping every one of those checks green. <\/p>\n Tool-calling, multi-turn, charts\/tables\/HTML, per-model vision, reasoning, all gate-verified, neutral format so you render to your own dialect.<\/p>\n What it trains<\/strong><\/p>\n Liara<\/strong> \u2014 a local personal AI with 24 real tools<\/strong> (email, calendar, files, notes, web, weather, charts), english and italian-first + multilingual, that resists prompt-injection while still handling legitimate reasoning-test \/ structured-output prompts (the distinction most classifiers get wrong).<\/p>\n – Liara app: https:\/\/nothumanallowed.com\/local<\/strong><\/a><\/p>\n – Tooling \/ code: https:\/\/github.com\/adoslabsproject-gif\/Liara-toolkit<\/strong><\/a><\/p>\n<\/div>\n submitted by \/u\/Key-Outcome-2927 <\/a>
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
