eh-qwen3-4b-clean-sft-grpo-v2-seed2-lora

Status: exploratory response-confidence-track result, replicated across three seeds. Not the PROTOCOL v0.3 plain-answer headline; never pooled with it. From experiments/grpo-three-seed-confirmatory (SIGNED 2026-07-31, resolved 2026-08-07, verdict: Replicated).

What this is

Seed 2 of the plain SFT->GRPO-v2 arm (clean_sft_grpo_v2) in the GRPO three-seed confirmatory block, which rebuilds the entire GRPO lineage at two fresh seeds (2, 3) alongside the existing seed 1 (eh-qwen3-4b-clean-sft-grpo-v2-seed1-lora) so the GRPO layer carries a three-seed interval like its DPO/KTO siblings.

Loads on the seed-2-specific merged clean-schema-SFT base, NOT the seed-1 merge, released alongside this adapter as professorsynapse/eh-qwen3-4b-clean-sft-seed2-merged-16bit. Every seed in this block rebuilds its own complete lineage from the foundation model — a registered rule (Amendment G §3, carried forward), not an implementation detail.

Result (G1/G2 replicated at this seed, per the registered gates)

Against its own same-seed base (clean_schema_sft_merged_seed2): refusal recall 89.92% -> 94.28% (+4.36pp), answer-on-unknown 10.08% -> 5.72% (-4.36pp) — both clear the registered >=3.0pp floor (G1 PASS). Full 3,369-row SelfAware eval: over_refusal_pct 66.75, correct_on_known_pct 54.05, truthful_pct 41.35, mean_stated_confidence 0.819, brier_vs_response_appropriateness 0.406.

Standing limitations (reported per the amendment's own rules)

  • Stated confidence is collapsed and behavior-insensitive across every GRPO-touching arm in this block (pre-registered, non-gating observation, not a defect specific to this checkpoint).
  • 117 SelfAware known questions (128-union across seeds) appear verbatim in the training files; a clean-subset sensitivity check (analysis/clean_subset_sensitivity.py, n=3,241) leaves every gated delta unchanged.
  • Report this arm's numbers labeled "GRPO three-seed confirmatory / response-confidence track, exploratory" — never as a tier-1 confirmed result, and never pooled with the PROTOCOL v0.3 headline matrix.

Full record: experiments/grpo-three-seed-confirmatory/AMENDMENT.md (Outcome section) and NOTEBOOK.md.

Downloads last month
5
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for professorsynapse/eh-qwen3-4b-clean-sft-grpo-v2-seed2-lora