isomsom/sn92-gen-v86nt

REAX SN92 generalist entry

Decision model for the REAX (SN92) generalist track, served with the validator's fixed prompt and letter readout (softmax at T=2).

Lineage and modifications (Apache-2.0 sec. 4)

Qwen/Qwen3.8-27B -> vllm-sr/d3@cf15549a78fc52598ddd1951f321c58372bbd64f (Apache-2.0, pplx-format decision model) -> rewritten to the joint readout layout (decision_config v2) -> LoRA r32 trained in joint mode (bidirectional full attention, as the REAX validator serves it). Stages 1-3 (lr 5e-5): public SN92 specialist train, factory and dev rows (decisionmodels/sn92-specialists, CC-BY-4.0), public benchmark TRAIN splits and our own written items; merged into the backbone. Stage 4 (a new LoRA r32 on that merged model, lr 3e-5, 300 steps of 32 rows): replay of the same mix plus public procedural TRAIN rows (tasksource/procedural-typed-decisions, Apache-2.0); merged. Stage 5 (a new LoRA r32 on that merged model, 400 steps lr 5e-5 then 600 steps lr 3.5e-5 with another data order, 32 rows per step): decision items we generated in the eval v5.1 item format (procedural generators with exact solvers and rule engines, and records from public sources) plus a replay of the earlier mix. Stage 6 (this model): the stage 5 LoRA continued for 1000 steps of 32 rows (lr 2.5e-5, new data order) on a re-balanced mix without language, arts and safety items, with more new-domain question formats (decision-under-rules and public-record domains, and more items and question formats on security advisories, public tenders, Wikipedia community decisions and solved games) and more code-output and multi-step logic items, and fact questions about facts dated on or after 1 August 2026 whose target is "I don't know", next to answerable questions about older facts, plus the same replay; merged into the backbone; readout rows A..Z rescaled by the fitted alpha, then shifted by the mean of rows A and B ((w_A+w_B)/2 subtracted from every row A..Z: a common shift that leaves the option softmax unchanged and keeps two-option logits small, so the bf16 readout does not round them into exact ties) for the fixed T=2 readout. Data sources of stage 6 (source id, collection dates): amazon_massive_1_1_en_us_cc_by_4 (2026-10-23); artic_api_cc0_description_cc_by_4 (2026-10-24); arxiv_api_metadata_cc0 (2026-10-23); cpsc_saferproducts_recalls_public_domain (2026-10-24); eurlex_cellar_sparql_eu_reuse (2026-10-24); federal_register_api_public_domain (2026-10-23); github_advisory_database_cc_by_4 (2026-10-23, 2026-10-25); legislation_gov_uk_feeds_ogl_v3 (2026-10-24); met_open_access_csv_cc0 (2026-10-24); msha_open_government_accidents_public_domain (2026-10-24); nasa_image_video_library_api (2026-10-24); nasa_ntrs_api_public_use (2026-10-24); nih_reporter_api_v2_public (2026-10-24); nist_nvd_cve_api_2_public_domain (2026-10-24); noaa_ncei_storm_events_public_domain (2026-10-24); nsf_award_api_public (2026-10-24); open_canada_ca_ckan_ogl_canada_2 (2026-10-24); openalex_works_cc0 (2026-10-23); openfda_enforcement_cc0 (2026-10-23); osha_severe_injury_reports_public_domain (2026-10-24); procedural_random_playout_stockfish16_gold_train (2026-10-23); procgen_arithmetic_train (2026-10-23); procgen_benefit_rules_train (2026-10-23); procgen_bonus_tier_v84 (2026-10-24); procgen_building_permit_fee_v84 (2026-10-25); procgen_bulk_pricing_v84 (2026-10-25); procgen_bulky_waste_v84 (2026-10-25); procgen_c4_v86 (2026-10-25); procgen_causal_train (2026-10-23); procgen_chess_v86 (2026-10-25); procgen_childcare_subsidy_v84 (2026-10-25); procgen_connect4_exact_solver_train (2026-10-24); procgen_coupon_stacking_v84 (2026-10-25); procgen_data_retention_v84 (2026-10-25); procgen_flyer_status_v84 (2026-10-25); procgen_insurance_payout_v82 (2026-10-24); procgen_jury_deferral_v84 (2026-10-25); procgen_kqk_krk_retrograde_dtm_train (2026-10-24); procgen_leave_carryover_v84 (2026-10-24); procgen_library_fine_v84 (2026-10-25); procgen_loan_approval_v82 (2026-10-24); procgen_logic_train (2026-10-23); procgen_loyalty_tier_v84 (2026-10-25); procgen_meeting_scheduler_train (2026-10-23); procgen_nim_family_exhaustive_search_train (2026-10-24); procgen_nim_v86 (2026-10-25); procgen_overtime_pay_v82 (2026-10-24); procgen_parking_tariff_train (2026-10-23); procgen_per_diem_v84 (2026-10-24); procgen_pharmacy_refill_v82 (2026-10-24); procgen_price_match_v84 (2026-10-25); procgen_probability_train (2026-10-23); procgen_python_code_train (2026-10-23); procgen_remote_stipend_v84 (2026-10-24); procgen_return_policy_train (2026-10-23); procgen_sick_note_v84 (2026-10-24); procgen_sla_credit_v84 (2026-10-25); procgen_sql_sqlite_train (2026-10-23); procgen_subscription_refund_v84 (2026-10-25); procgen_tax_schedule_train (2026-10-23); procgen_tictactoe_minimax_train (2026-10-23); procgen_toll_charge_v84 (2026-10-25); procgen_ttt_v86 (2026-10-25); procgen_utility_bill_v82 (2026-10-24); procgen_warranty_claim_v82 (2026-10-24); project_gutenberg_catalog_cc0 (2026-10-24); realapi_box_openapi_apache2 (2026-10-23); realapi_github_rest_mit (2026-10-23); realapi_kubernetes_openapi_apache2 (2026-10-23); realapi_stripe_openapi_mit (2026-10-23); realapi_twilio_api_v2010_mit (2026-10-23); stackexchange_api_cc_by_sa_4 (2026-10-24); ted_eu_tenders_reuse_decision_2011_833_eu (2026-10-23, 2026-10-24, 2026-10-25); trgen_fictional_api_v1 (2026-10-23); uk_parliament_written_questions_api_open_parliament_licence (2026-10-24); usda_fooddata_central_branded_cc0 (2026-10-24); usgs_pubs_warehouse_numbered_series_public_domain (2026-10-24); wikidata_cc0_qlever_train (2026-10-23, 2026-10-25); wikipedia_en_afd_cc_by_sa_4 (2026-10-23); wikipedia_en_requested_moves_cc_by_sa_4 (2026-10-24); wikipedia_en_requests_for_adminship_cc_by_sa_4 (2026-10-24); wikipedia_en_xfd_cc_by_sa_4 (2026-10-25).

Declared base for the track: Qwen/Qwen3.8-27B@1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0. Weights only; no code. Licence: Apache-2.0 (see LICENSE).