{
"$type": "site.standard.document",
"bskyPostRef": {
"cid": "bafyreibegewcl7ihelxtndylzrzzpucwtnpxxmuxp4rvks7g3malzvp66i",
"uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mgsqa2blgwd2"
},
"path": "/t/zero-forgetting-across-4-benchmarks-on-mistral-7b-interactive-results-dashboard/174199#post_1",
"publishedAt": "2026-03-11T19:12:19.000Z",
"site": "https://discuss.huggingface.co",
"tags": [
"huggingface.co",
"Zero Forgetting Benchmarks - a Hugging Face Space by Fourwheels2512",
"https://mhc-finetune-saas-zrtokzlkbnue9zsk7jfgad.streamlit.app"
],
"textContent": "We’ve been working on continual learning for LLM fine-tuning — training one model sequentially across\nmultiple domains without catastrophic forgetting. After 6 months of R&D and 50+ failed experiments (EWC,\nreplay, KD, gradient projection), we have a method that works.\n\n4 independent benchmarks on Mistral-7B:\n\n * Research benchmark (5 domains, 3 seeds) — -0.17% drift vs +43% forgetting with naive LoRA\n * Walmart enterprise (4 domains) — BERTScores 0.82–0.94 across all domains retained\n * Salesforce enterprise (5 domains) — Positive backward transfer: retention BERTScores improved with\neach new domain (0.889 → 0.907)\n * Dental stress test (8 domains, 2 seeds) — Gradient norms stable throughout, zero crashes\n\n\n\nSpectral norm locked at 1.0 across every experiment. Standard LoRA crashed at step 43 with gradient norm\n263. Ours: peak under 6. No replay buffers, no EWC, no knowledge distillation.\n\nThe adapter is ~0.1% additional parameters, works with any LoRA/QLoRA setup.\n\nInteractive benchmark dashboard with charts:\n\nhuggingface.co\n\n### Zero Forgetting Benchmarks - a Hugging Face Space by Fourwheels2512\n\nZero forgetting in LLM fine-tuning — 4 benchmarks\n\nLive product (free tier, no credit card): https://mhc-finetune-saas-zrtokzlkbnue9zsk7jfgad.streamlit.app\n\nUS patent pending. Would love to hear from anyone working on continual learning or dealing with\nforgetting in multi-domain fine-tuning.",
"title": "Zero Forgetting Across 4 Benchmarks on Mistral-7B — Interactive Results Dashboard"
}