{"slug": "training-a-57m-edge-model-sft-tool-calling-done-now-branching-into-3-rlvs-loops", "title": "Training a 57M Edge Model: SFT Tool-Calling Done, Now Branching into 3 RLVS Loops", "summary": "A developer trained Nova, a 57.2M-parameter edge model, through supervised fine-tuning with a syntactically verified tool-calling dataset and Program-of-Thought execution traces, reaching 100% deterministic tool-calling precision. Rather than continue single-track training, the developer cloned the base SFT checkpoint into three forks, each trained under a specialized Reinforcement Learning with Verifiable Supervision verifier covering cybersecurity, ARC-AGI, and balanced desktop execution, to avoid gradient interference between skills.", "body_md": "**TL;DR:** We just completed Supervised Fine-Tuning (SFT) on Nova, our home-grown 57M parameter edge model, hitting 100% deterministic tool-calling precision. Instead of continuing single-track training, we are cloning the base weights into three distinct forks, each trained under a specialized Reinforcement Learning with Verifiable Supervision (RLVS) verifier: Cybersecurity, ARC-AGI, and Balanced Desktop Execution.\n\nWhen training a sub-100M parameter model (Nova is exactly 57.2M parameters), the biggest danger is **format drift** during structured JSON tool calls. Traditional LLMs often hallucinate brackets or invent parameter keys when parameters drop below 1B.\n\nBy curating a clean, syntactically-verified tool-calling dataset with Program-of-Thought (PoT) execution traces:\n\nWith the base grammar and tool execution locked in, we reached a crossroads: *How do you teach a 57M model high-level reasoning without bloating its parameter count?*\n\nInstead of forcing a single 57M model to master everything and diluting its weights, we are taking the **identical base SFT checkpoint** and cloning it into three parallel training tracks:\n\n                ┌─────────────────────────────────┐\n\n                  │    Nova 57M Base SFT Checkpoint │\n\n                  │   (100% Deterministic Tools)    │\n\n                  └────────────────┬────────────────┘\n\n                                   │ Clone Weights\n\n    ┌──────────────────────────────┼──────────────────────────────┐\n\n    ▼                              ▼                              ▼\n\n┌────────────────────────┐ ┌────────────────────────┐ ┌────────────────────────┐ │ Track 1: Nova-Sec │ │ Track 2: Nova-ARC │ │ Track 3: Nova-Balanced│ │ RLVS Focus: │ │ RLVS Focus: │ │ RLVS Focus: │ │ • Prompt injection │ │ • ARC-AGI visual grid │ │ • Low-latency Cowork │ │ • Secure code audit │ │ • Spatial transform │ │ • General debugging │ │ • Exploit verification│ │ • Step-wise logic │ │ • Real-time OS agent │ └────────────────────────┘ └────────────────────────┘ └────────────────────────┘\n\nIn typical supervised fine-tuning, training on diverse datasets causes **gradient interference** (learning security rules degrades spatial reasoning, and vice-versa). \n\nBy freezing the SFT tool-calling base and training three separate RLVS specialists:\n\nFor those training small-language models (SLMs) under 100M parameters:\n\nDrop your thoughts below!", "url": "https://wpnews.pro/news/training-a-57m-edge-model-sft-tool-calling-done-now-branching-into-3-rlvs-loops", "canonical_source": "https://dev.to/achyut_srivastava_2a77b5d/-training-a-57m-edge-model-sft-tool-calling-done-now-branching-into-3-rlvs-loops-1k24", "published_at": "2026-11-20 11:11:53+00:00", "updated_at": "2026-11-20 11:14:04.947839+00:00", "lang": "en", "topics": ["machine-learning", "large-language-models", "ai-agents", "ai-research", "ai-tools"], "entities": ["Nova", "ARC-AGI"], "also_reported_by": [], "takeaway": "Cloning one SFT checkpoint into separate RLVS specialist forks avoids gradient interference when teaching a sub-100M model multiple skills, while a syntactically verified tool-calling dataset with Program-of-Thought traces locks in deterministic JSON tool calls.", "recommendation": {"advice": "Freeze a verified tool-calling SFT base and train separate RLVS specialist forks per skill instead of one mixed dataset, to prevent gradient interference in sub-100M models.", "playbook": "training-data", "playbook_url": "https://wpnews.pro/playbooks/training-data"}, "alternates": {"html": "https://wpnews.pro/news/training-a-57m-edge-model-sft-tool-calling-done-now-branching-into-3-rlvs-loops", "markdown": "https://wpnews.pro/news/training-a-57m-edge-model-sft-tool-calling-done-now-branching-into-3-rlvs-loops.md", "text": "https://wpnews.pro/news/training-a-57m-edge-model-sft-tool-calling-done-now-branching-into-3-rlvs-loops.txt", "jsonld": "https://wpnews.pro/news/training-a-57m-edge-model-sft-tool-calling-done-now-branching-into-3-rlvs-loops.jsonld"}}