Works with the AI Agents you already use
Post-Training Methods Pack: SFT, LoRA, DPO, GRPO Setup
Picks SFT, DPO-family, or RL for your goal and sets it up with starting hyperparameters and the failure signals to watch.
$5
Post-Training Methods Pack: SFT, LoRA, DPO, GRPO Setup
Example session with this skill installed
We ran DPO on our SFT model with 5k preference pairs. Win rate went up but answers got way longer and the model now rambles on simple questions. lr 5e-5 full fine-tune, beta 0.1, 3 epochs.
- Read your context and instructions
- Compiled the post-training methods
Most likely cause: the learning rate is about 10x too high for a full DPO fine-tune, and 3 epochs overfits the pairs; together they amplify any length gap in your data. Check first: average chosen vs rejected length - if chosen is much longer, the model learned "longer wins." Fix: lr ~5e-7 to 1e-6, 1 epoch, rebalance or length-match pairs, and log chosen and rejected log-probs (both falling = drift). Gate: win rate holds on a length-controlled eval with no drop on short-answer slices. Next action: compute the length gap in the pairs.
Connects securely to your tools. The creator never sees your data.
About this skill
"Should we do SFT, DPO, or RL?" is the most expensive question in fine-tuning to get wrong. This skill makes your agent act as the training lead: it picks the simplest method that can work and sets it up properly.
What it does
- Decision path from prompting/RAG → SFT/LoRA → DPO/IPO/KTO/ORPO → RLVR with GRPO → PPO-style RLHF → distillation, with a cost and risk table.
- Starting hyperparameters for LoRA/QLoRA (rank, alpha, target layers, learning rate, epochs), DPO (beta, learning rate, epochs), and GRPO (group size, KL, prompt difficulty filtering).
- The failure signals for each method: template and masking bugs, forgetting, likelihood displacement in DPO, length runaway and reward hacking in RL.
- Verifier-backed training for math and code reasoning.
- Debug mode: share a config, loss/reward curves, or bad outputs and get the most likely cause first.
Part of the Post-Training series. The full Post-Training Gatekeeper bundles all four packs.
Advises on methods and configs; it does not run training jobs. Hyperparameters are starting points to sweep, not tuned values for your model.
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