# Training a 57M Edge Model: SFT Tool-Calling Done, Now Branching into 3 RLVS Loops

> Source: <https://dev.to/achyut_srivastava_2a77b5d/-training-a-57m-edge-model-sft-tool-calling-done-now-branching-into-3-rlvs-loops-1k24>
> Published: 2026-11-20 11:11:53+00:00

**Recommendation:** 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.
More in the playbook "Fine-tuning and data": <https://wpnews.pro/playbooks/training-data>

**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.

When 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.

By curating a clean, syntactically-verified tool-calling dataset with Program-of-Thought (PoT) execution traces:

With 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?*

Instead 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:

                ┌─────────────────────────────────┐

                  │    Nova 57M Base SFT Checkpoint │

                  │   (100% Deterministic Tools)    │

                  └────────────────┬────────────────┘

                                   │ Clone Weights

    ┌──────────────────────────────┼──────────────────────────────┐

    ▼                              ▼                              ▼

┌────────────────────────┐ ┌────────────────────────┐ ┌────────────────────────┐ │ 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 │ └────────────────────────┘ └────────────────────────┘ └────────────────────────┘

In typical supervised fine-tuning, training on diverse datasets causes **gradient interference** (learning security rules degrades spatial reasoning, and vice-versa). 

By freezing the SFT tool-calling base and training three separate RLVS specialists:

For those training small-language models (SLMs) under 100M parameters:

Drop your thoughts below!
