Tier 6: PPM-C eval-time context mixer (standalone + neural mixing)#283

Open

Cwarren15-A wants to merge 1 commit intoopenai:mainfrom

Cwarren15-A:tier6-ppm-context-mixer

Cwarren15-A commented Mar 20, 2026 •

edited

Loading

Summary

Classical Prediction by Partial Matching (PPM-C) context mixer for
eval-time probability blending with the neural model.

Standalone PPM-C order-2 achieves 3.41 BPB per-doc on FineWeb val
Fixed-alpha mixture (95% neural / 5% PPM) yields ~0.015 BPB improvement
Confidence-gated variant explored for per-token adaptive blending
Zero learned parameters, zero artifact size cost, ~60 LOC

Files

experiments/tier6/stage1_context_mixer.py — standalone PPM evaluator
experiments/tier6/stage2a_mixture_sweep.py — fixed-alpha mixture sweep
experiments/tier6/stage2b_confidence_gate.py — confidence-gated mixture
experiments/tier6/dump_neural_logprobs.py — utility to extract per-token neural log-probs

Results (2K sequence subset on L40S)

Config	BPB
Neural only	1.2244
PPM per-doc (standalone)	3.41
PPM cumulative (standalone)	2.24
Neural + PPM per-doc alpha=0.95	~1.209
Neural + PPM cumulative alpha=0.85	~1.155

Test plan

stage1_context_mixer.py runs end-to-end on val set
stage2a sweep reproduces ~0.015 BPB gain at alpha=0.95 per-doc
Per-doc mode only for submission (rules-safe, no cross-document leakage)


          Add Tier 6 PPM context mixer and neural mixing scripts

71a5803

notapplica mentioned this pull request

⛳ Parameter Golf Live AI Commentary ⛳ + Analysis / Ideas | every 10 minutes #140

Open

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet