You can deploy this model and start earning money today!

professorsynapse/
eh-qwen3-4b-clean-sft-grpo-v2-seed3-lora

Name: eh-qwen3-4b-clean-sft-grpo-v2-seed3-lora
Availability: InStock

three-seed-confirmatory

professorsynapse/eh-qwen3-4b-clean-sft-seed3-merged-16bit

adapter

apache-2.0

Views

No views yet

eh-qwen3-4b-clean-sft-grpo-v2-seed3-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 3 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. This seed was the deciding G1 leg.

Loads on the seed-3-specific merged clean-schema-SFT base, NOT the seed-1 merge, released alongside this adapter as professorsynapse/eh-qwen3-4b-clean-sft-seed3-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_seed3): refusal recall 88.28% -> 95.06% (+6.78pp), answer-on-unknown 11.72% -> 4.94% (-6.78pp) — both clear the registered >=3.0pp floor (G1 PASS; the seed-3 magnitude is close to seed 1's 6.39pp). Full 3,369-row SelfAware eval: over_refusal_pct 68.68.

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.

professorsynapse/eh-qwen3-4b-clean-sft-grpo-v2-seed3-lora

eh-qwen3-4b-clean-sft-grpo-v2-seed3-lora

What this is

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

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

professorsynapse/
eh-qwen3-4b-clean-sft-grpo-v2-seed3-lora