{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreie4ciwmetcm6udtp7vjdzw7myz3mkidwq5dybnvygkbsxjkgzu3l4",
    "uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mfsmbfyeftn2"
  },
  "path": "/t/crma-stable-fine-tuning-continual-learning-for-small-llms/173817#post_1",
  "publishedAt": "2026-02-27T01:06:41.000Z",
  "site": "https://discuss.huggingface.co",
  "tags": [
    "https://fourwheels2512--crma-finetune-fastapi-app.modal.run"
  ],
  "textContent": "CRMA: Stable Fine-Tuning + Continual Learning for Small LLMs\n\nWe’ve been building CRMA (Constrained Residual Mixing Adapter) — a small adapter that attaches to every layer of a\nlanguage model during fine-tuning. It applies a mathematical constraint that keeps training stable: the model can\nlearn new information but can’t overwrite what it already knows.\n\nInspired by “mHC: Manifold-Constrained Hyper-Connections” (arXiv:2512.24880) by Zhenda Xie, Yixuan Wei, et al. — not\nequivalent.\n\nWhat it does — two capabilities:\n\n  1. Fine-tuning stability\n\n\n  * Peak gradient norm reduced 39–84% vs standard LoRA\n  * Near-identity initialization — no cold-start collapse\n  * Works with QLoRA (4-bit) on TinyLlama-1.1B, Mistral-7B, Gemma-2B\n  * All stability claims are empirically measured per run, not theoretical\n\n\n  2. Continual learning\n\n\n  * Train sequentially on multiple domains — medical, legal, code, finance\n  * -0.1% backbone drift across 4 domains (vs +351% catastrophic forgetting with naive sequential training)\n  * Each domain gets its own adapter; the shared backbone stays stable\n  * No replay buffers, no growing memory — swap adapters at inference\n\n\n\nMeasured on:\n\n  * TinyLlama-1.1B-Chat (1.1B params, Apache 2.0)\n  * Mistral-7B-v0.3 (7B params, Apache 2.0)\n  * Modal A10G GPU\n\n\n\nTry it:\n\n  * API: https://fourwheels2512--crma-finetune-fastapi-app.modal.run\n  * Free tier: 3 runs/day on TinyLlama, no credit card needed\n  * Pro: pay-as-you-go credits, starting at $5\n\n\n\nThe fine-tuning API is live. Continual learning is available via the /start_cl_run endpoint — bring a base fine-tuned\nrun and add new domains without losing previous ones.\n\nBuilt by Kiran Nayudu. Feedback welcome.",
  "title": "CRMA: Stable Fine-Tuning + Continual Learning for Small LLMs"
}