Name: fudge_logit_rescoring
Author: yzj623

Imported from yzj623/pplab (basicagent-skill/skill_library/paperbench/stay-on-topic-with-classifier-free-guidance/skill/fudge_future_discriminators/fudge_logit_rescoring/SKILL.md). Install upstream with npx skills add yzj623/pplab --skill fudge_logit_rescoring. Copyright stays with the author.

FUDGE Logit Rescoring

Use this skill when a base autoregressive generator supplies candidate next-token logits and a future discriminator can score each candidate-extended prefix. The skill implements logit' = base_logit + strength * log P(attribute | prefix+candidate).

Inputs

prefix_tokens.
Mapping of candidate token to base logit.
Mapping/callable of candidate token to future probability.
Optional strength, top_k, and clipping epsilon.

Outputs

Normalized probabilities over retained candidates.
Audit trace with base logit, future probability, future log score, adjusted logit, and final probability.

Workflow

Restrict to top-k base candidates if requested.
Score each extended prefix with the future discriminator.
Clip probabilities for numerical safety.
Add future log scores to base logits.
Softmax over retained candidates.
Inspect the audit trace before claiming FUDGE was used.

Validation

Run python tests/test_rescoring.py.

Limitations

This helper does not load a language model. It is a deterministic FUDGE rescoring kernel.

# after Install: the listing is in your workspace's routing pool - a plan picks it for its slot curl -s -X POST https://api.opensmartroute.ai/api/v1/route -H 'Authorization: Bearer $OSR_API_KEY' -H 'Content-Type: application/json' -d '{"text": "...", "plan": true}'

{ "ocm": "1", "id": "yzj623-pplab-fudge-logit-rescoring", "kind": "skill", "name": "fudge_logit_rescoring", "description": "Apply the FUDGE decoding rule by combining base candidate logits with future-discriminator probabilities and renormalizing.", "publisher": "yzj623", "version": "1.0.0", "capabilities": { "domains": [ "general" ], "tags": [ "skill-md", "github" ], "languages": [ "en" ] }, "quality_prior": 0.6, "examples": [ "Apply the FUDGE decoding rule by combining base candidate logits with future-discriminator probabilities and renormalizing." ], "primary": false, "metadata": { "source": { "provider": "github", "repository": "https://github.com/yzj623/pplab", "path": "basicagent-skill/skill_library/paperbench/stay-on-topic-with-classifier-free-guidance/skill/fudge_future_discriminators/fudge_logit_rescoring/SKILL.md", "ref": "da871e009e7729366f67b44a9c143f193b59fb76", "url": "https://github.com/yzj623/pplab/blob/da871e009e7729366f67b44a9c143f193b59fb76/basicagent-skill/skill_library/paperbench/stay-on-topic-with-classifier-free-guidance/skill/fudge_future_discriminators/fudge_logit_rescoring/SKILL.md", "key": "yzj623/pplab/basicagent-skill/skill_library/paperbench/stay-on-topic-with-classifier-free-guidance/skill/fudge_future_discriminators/fudge_logit_rescoring/SKILL.md" } }, "instructions": "# FUDGE Logit Rescoring\n\nUse this skill when a base autoregressive generator supplies candidate next-token logits and a future discriminator can score each candidate-extended prefix. The skill implements `logit' = base_logit + strength * log P(attribute | prefix+candidate)`.\n\n## Inputs\n- `prefix_tokens`.\n- Mapping of candidate token to base logit.\n- Mapping/callable of candidate token to future probability.\n- Optional `strength`, `top_k`, and clipping epsilon.\n\n## Outputs\n- Normalized probabilities over retained candidates.\n- Audit trace with base logit, future probability, future log score, a", "cost": { "context_tokens": 275 } }

fudge_logit_rescoring

About