DEDUPED Hot AI repos, turned into use cases Get 5 a day

How to test soft-target training with RLCDLoss on your own labels with Open Jev

Use probability distributions as targets and run one loss and backward step to study the calibration-oriented objective.

  • Research
  • Advanced
  • 60 minutes
  • From the README
  • Any agent

Who it is forA researcher studying calibration who wants to see how soft probability targets flow through the RLCDLoss objective.

The problemStandard one-hot labels force certainty, and the researcher needs a way to keep disagreement visible in training targets.

The outcomeA runnable loss computation over custom probability targets, with gradients produced by backward().

What you need first

  • Python 3.10 or newer (recommended in the README)
  • git installed
  • PyTorch 2.0 or newer
  • Basic familiarity with running Python scripts from a terminal

Steps

  1. Install the repo and dependency as in the install section: git clone https://github.com/kyegomez/open-jev then python -m pip install "torch>=2.0"
  2. Create one target tensor per question, where each row is a probability distribution rather than a one-hot label
  3. Instantiate the loss with RLCDLoss() and call it with the model, states, questions, and targets
  4. Run loss.backward() on the returned loss
  5. Optionally pass paraphrased or key-shuffled versions of the same state through augmented_states to test consistency (inferred, check the README)

Prompt for your agent

In the repo https://github.com/kyegomez/open-jev, build soft probability targets for <your question set> and compute RLCDLoss on the model with one backward pass. Expected output is a scalar loss tensor with gradients; verify that each target row sums to 1 before calling the loss.

Caveat. The training recipe is described as a hypothesis, the repo has no trained checkpoints, and a single step on random weights does not show that the model becomes calibrated.

Share: X LinkedIn Hacker News Email Copy link

Disclaimer

Deduped lists open-source repositories that belong to other people. We do not own, maintain, audit, or vouch for any of them, and a repo appearing here is not an endorsement or a safety check. The use cases, steps, and prompts are written by AI and may be incomplete, outdated, or just wrong. Where a use case is marked "Seen in the wild", we link to public posts where someone described something similar, but we have not verified those posts and the use case itself is still our AI's own write-up. Use all of this at your own risk. Only run code, commands, or agents on systems and accounts you are authorized to use, and read the code and every command before you run it, especially anything that touches credentials, files, or the network. Nothing here is professional, security, legal, or financial advice. All trademarks and project names belong to their respective owners, and their use here does not imply affiliation. Content is provided as is, without warranties of any kind. Found a mistake, a safety problem, or something you want removed? Email [email protected] and we will look at it promptly. Full disclaimer page