Love seeing the Esper line keep shipping in the open. Curious how Esper 4 holds its general capabilities after stacking Titanium + Mitakihara + Tachibana — specializing that hard is usually where sequential fine-tunes start drifting on everything else. That retention problem is what we've been heads-down on at ModelBrew (modelbrew.ai); the multi-set diversity you're using is a real part of what helps. Grabbing these, thanks for open-sourcing.
Kiran N PRO
Fourwheels2512
AI & ML interests
None yet
Recent Activity
commentedon a paper 3 days ago
OpenUnlearning: Accelerating LLM Unlearning via Unified Benchmarking of
Methods and Metrics commentedon a paper 3 days ago
GUARD: Guided Unlearning and Retention via Data Attribution for Large Language Models commentedon a paper 3 days ago
Mechanistic Analysis of Catastrophic Forgetting in Large Language Models During Continual Fine-tuningOrganizations
posted an update 3 months ago
Post
882
We trained Mistral 7B, Qwen 8B, Gemma 9B models on 5 domains sequentially to test catastrophic forgetting.
We achieved zero forgetting with medical knowledge retained at 100% after adding enterprise, finance, military, and real estate domains on top.
Most fine-tuned models catastrophically forget everything they learned when you train them on something new. We built a continual learning engine that prevents this. First of its kind.
We're shipping it as a SaaS platform at modelbrew.ai - dataset optimization + fine-tuning + continual learning in one pipeline.
I'm looking for ML fine-tuning engineers and researchers who want to test this. DM me or comment below.
We achieved zero forgetting with medical knowledge retained at 100% after adding enterprise, finance, military, and real estate domains on top.
Most fine-tuned models catastrophically forget everything they learned when you train them on something new. We built a continual learning engine that prevents this. First of its kind.
We're shipping it as a SaaS platform at modelbrew.ai - dataset optimization + fine-tuning + continual learning in one pipeline.
I'm looking for ML fine-tuning engineers and researchers who want to test this. DM me or comment below.
posted an update 3 months ago
Post
959
Dataset Optimizer (free) + fine tuning + continual learning platform with zero forgetting.
try it at modelbrew.ai
try it at modelbrew.ai
posted an update 4 months ago
Post
148
Zero Forgetting in LLM Fine-Tuning — 4 Benchmarks, All Domains Retained
We tested sequential fine-tuning on Mistral-7B across 4 independent benchmarks (5, 4, 5, and 8 domains).
Standard LoRA forgets 38–49% of prior knowledge per domain. Our continual learning adapter: -0.17% drift.
The Salesforce 5-domain test showed positive backward transfer — the model got better at old domains as it learned new ones (retention BERTScore: 0.889 → 0.907).
No replay buffers. No EWC. No knowledge distillation. Spectral norm locked at 1.0. Naive LoRA crashed at gradient norm 263. Ours: under 6.
Interactive benchmark dashboard: Fourwheels2512/zero-forgetting-benchmarks
Live product (free tier): https://mhc-finetune-saas-zrtokzlkbnue9zsk7jfgad.streamlit.app
Patent pending. 7 technical reports. 196 automated tests. Solo founder, 6 months of R&D.
We tested sequential fine-tuning on Mistral-7B across 4 independent benchmarks (5, 4, 5, and 8 domains).
Standard LoRA forgets 38–49% of prior knowledge per domain. Our continual learning adapter: -0.17% drift.
The Salesforce 5-domain test showed positive backward transfer — the model got better at old domains as it learned new ones (retention BERTScore: 0.889 → 0.907).
No replay buffers. No EWC. No knowledge distillation. Spectral norm locked at 1.0. Naive LoRA crashed at gradient norm 263. Ours: under 6.
Interactive benchmark dashboard: Fourwheels2512/zero-forgetting-benchmarks
Live product (free tier): https://mhc-finetune-saas-zrtokzlkbnue9zsk7jfgad.streamlit.app
Patent pending. 7 technical reports. 196 automated tests. Solo founder, 6 months of R&D.