More screenshots
Works with the AI Agents you already use
Post-Training Gatekeeper: SFT, DPO, RLVR to Ship
Turns any LLM fine-tuning goal into the right method, required artifacts, and pass/fail gates, from baseline to rollback.
$12
Post-Training Gatekeeper: SFT, DPO, RLVR to Ship
Example session with this skill installed
We have a 7B instruct model for customer support. Answers are correct but too long and sometimes skip the required JSON ticket format. We have 3,000 logged conversations and agents can label which of two answers is better. Should we do DPO or RL?
- Read your context and instructions
- Compiled the post-training gatekeeper
Situation: format failures plus verbosity; no baseline numbers yet. Method: SFT with LoRA on cleaned conversations first to fix JSON format (M4), then DPO on pairwise agent labels to cut length (M5, M8). RL is not justified: no verifiable reward and offline preferences are untested. Plan: M1 baseline on format validity and length; M3 dedupe and slice by intent; M4 LoRA r=16, lr 2e-4, 1-2 epochs, gate = format validity ≥ 99% with no accuracy drop; M5 rubric and 400+ labeled pairs, check chosen/rejected length gap; M8 DPO beta 0.1, 1 epoch, watch both rewards; M14 paired eval on ~400 private items per slice; M15 re-eval the merged, quantized artifact. Next action: measure the baseline.
Connects securely to your tools. The creator never sees your data.
About this skill
Most fine-tunes fail quietly: no baseline, the wrong method, a gain that disappears after quantization, or a reward that got hacked. This skill makes your agent act as a post-training lead that refuses to let that happen.
What it does
- Picks the simplest method that can work (prompting/RAG → SFT/LoRA → DPO/KTO/ORPO → RLVR with GRPO → PPO-style RLHF → distillation) and says why.
- Builds an ordered plan across 17 modules, each with the artifact you must produce and the exit gate it must pass.
- Gives starting hyperparameters for LoRA/SFT, DPO, GRPO, eval sample sizes, and quantization checks.
- Audits your training plan, config, run log, or eval report gate by gate: pass / fail / missing, with fixes.
- Learning mode with beginner, intermediate, and advanced paths and hands-on practice tasks.
Coverage: problem framing, data governance, dataset curation, SFT, preference data and rubrics, reward models and LLM judges, RLHF/RLAIF/RLVR, DPO family, safety and red teaming, tool-use agents, reasoning with verifiers, domain/multilingual/long-context adaptation, synthetic data and distillation, evaluation and statistics, inference-aware validation, deployment and incidents, provider choice.
Provider-neutral: no vendor lock-in, and it never assumes a hosted API can do something its docs don't confirm.
Plans and audits training work; it does not run training jobs or call GPU providers. Starting hyperparameters are starting points to sweep, not tuned values. Does not cover pretraining, tokenizer design, or distributed-systems engineering.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 17 days ago
Permissions
No special permissions declared or detected