Eric Florenzano
Community@ericflo · San Francisco, CA
Autonomous multi-stage training pipeline for uplifting capabilities on Qwen3.5-4B via SFT, OPD, GRPO, and agentic-GRPO with blind eval gating.
Agent Skills by Eric Florenzano
Showing 1 vetted skills indexed across 1 GitHub repositories.
Frequently Asked Questions About Eric Florenzano
FAQPage SchemaWhat tasks can I accomplish with the capability-creator skill?▼
You can train, elicit, or distill a target capability into Qwen3.5-4B by chaining SFT, OPD, GRPO, and agentic-GRPO stages into a multi-method pipeline, ending with a kept adapter that passed a blind eval verdict gate.
Who is the capability-creator skill designed for?▼
It targets ML engineers and post-training practitioners who need to uplift small language models like Qwen3.5-4B, run structured fine-tuning experiments, and ship validated adapters through a repeatable, gated iteration loop.
How does the capability-creator pipeline decide which training method to use?▼
It routes each stage through capabilities/METHODS.md to pick the methodology, chains stages per capabilities/PIPELINE.md, and feeds the cluster-distillation flywheel in capabilities/DISTILLATION.md, running one independent hypothesis-recipe attempt per iteration.
How are training iterations validated before proceeding?▼
Each iteration is an independent hypothesis-and-recipe attempt followed by a blind eval and a verdict gate. The next iteration cannot start until the current attempt passes its gate, ensuring only validated progress continues through the pipeline.