Structure-RLCR for ReCCon (Alexps9yyy/rlcr4reccon)

Official model & dataset hub for Structure-RLCR, replacing all three probabilistic interfaces of the ReCCon multi-chain verification pipeline with proper-scoring GRPO calibrated Qwen3-8B and Qwen3-32B models.


1. Repository Layout

A. Best Trained Checkpoints (models/)

Subfolder Base Model ReCCon Stage Output Format & Objective Key OOD IFQA / Held-out Metric
models/RLCR-JOINT-8B-opt Qwen/Qwen3-8B Stage 1 (RLCR-Joint) <think><answer><analysis><confidence> (0.35 Acc + 0.65 Brier) IFQA Single-Chain Acc 83.19% (beats Native 35B/20B 81.64%), 3-Chain Acc 85.47%
models/RLCR-QA-8B-opt Qwen/Qwen3-8B Stage 3 (RLCR-QA) <think><analysis><belief> (Brier-only) v3-qa Held-out AUROC 79.38; IFQA Contra reccon 93.04 (85.1% false facts < 0.5)
models/RLCR-CRED-8B-opt Qwen/Qwen3-8B Stage 5 (RLCR-CRED) <think><analysis><belief> (Brier-only) IFQA Only-S5 p_rw 78.70 / Chain AUROC 73.99; Full-8B Div p_rw 79.36 ($p=0.0053$)
models/RLCR-JOINT-32B-opt Qwen/Qwen3-32B Stage 1 (RLCR-Joint) <think><answer><analysis><confidence> (0.35 Acc + 0.65 Brier) IFQA 3-Chain Mean AUROC 93.80, Acc 86.96%, cuts Stage 2–5 audit rate 30.7% -> 15.6%
models/RLCR-QA-32B-opt Qwen/Qwen3-32B Stage 3 (RLCR-QA) <think><analysis><belief> (Brier-only) v3-qa Held-out Acc 74.00%; IFQA 32B-Joint Contra (n=471) 92.21 -> 93.34
models/RLCR-CRED-32B-opt Qwen/Qwen3-32B Stage 5 (RLCR-CRED) <think><analysis><belief> (Brier-only) v5-cred Held-out AUROC 67.85; Full-32B IFQA AUROC 94.17 (s038) / Div 81.84 & Brier 9.81 (s034)

B. Training & Held-Out Datasets (data/) β€” Zero OOD IFQA Overlap

Dataset Path Stage Train / Val / Test Rows Description
data/v2-joint-ctx/hf_files/ Stage 1 (Joint) 11,455 / 197 / 197 HoVer + FactCHD + ExFEVER + SciFact statements with multi-query BM25 Wikipedia evidence (10% empty-evidence calibration)
data/v3-joint-cf/hf_files/ Stage 1 (Joint-CF) 15,251 / 197 / 197 Adds 3,796 zero-leakage ifqa_train counterfactual statements to v2-joint-ctx
data/ifqa-joint-ctx/hf_files/ Stage 1 OOD Eval β€” / β€” / 1,273 All 1,273 OOD IFQA test statements with pre-retrieved BM25 evidence for 3-chain Stage-1 inference
data/v2-qa/hf_files/ Stage 3 (QA base) 13,678 / 1,708 / 1,708 Base R_FACT_QA wrapped statements (ExFEVER, HoVer, ifqa_train, SciFact) with 6000-char BM25 evidence
data/v3-qa/hf_files/ Stage 3 (QA aligned) 15,858 / 1,234 / 1,234 Strictly excludes all IFQA rows and adds 2x balanced real Stage-3 (question, candidate, evidence) rows from HoVer + FactCHD verify.json
data/v4-qa-cf/hf_files/ Stage 3 (QA-CF) 23,402 / 1,708 / 1,708 Enriches v3-qa with 3,800 zero-leakage ifqa_train R_FACT_QA rows + 3,744 contrastive conditional Stage-3 candidate pairs (32.2% counterfactual, 0 overlap with OOD IFQA test)
data/v3-cred/hf_files/ Stage 5 (CRED base) 1,893 / 234 / 240 Per-chain meta-verification rows (y_i = 1[(p_i >= 0.5) == gold]) from HoVer + FactCHD
data/v4-cred-spread/hf_files/ Stage 5 (CRED spread) 1,926 / 234 / 240 Spread-only (max p_i - min p_i > 0.25), 2x FactCHD upsampled, 50% p_true parenthetical dropout
data/v8-cred-opt/hf_files/ Stage 5 (CRED opt) 3,852 / 234 / 240 Optimal clean Stage-5 pool combining 32B-QA + native 35B verify.json (spread-only on train, no rules, 15% p_true dropout)

C. ReCCon Evaluation Artifacts (eval_runs/ & benchmarks/)

  • eval_runs/e1_native/run/ifqa/{statements,chains,claims,verify,layer2}.json: Native ReCCon Stage 1–4 artifacts on the 1,262 OOD IFQA benchmark statements (387 divergent, 194 contradiction-only).
  • eval_runs/e1e2e3/run/ifqa/judge.json: Native Qwen3.6-35B-A3B Stage-5 reference judge scores (79.08 divergent AUROC / 93.72 full AUROC at mix=0.05).
  • eval_runs/cred_pool/{hover,factchd}/{statements,chains,claims,verify,layer2}.json: HoVer (1,000) & FactCHD (971) Stage 1–4 pools used to rescore Stage-3 candidates and reconstruct Stage-5 training sets.
  • benchmarks/{ifqa,hover,factchd,futurex,ragtruth_marco,truthfulqa}.json: Standalone evaluation statement sets.

2. Quick Download

# Download all training & evaluation datasets into a BTProp checkout
hf download Alexps9yyy/rlcr4reccon --include "data/*" --include "eval_runs/*" --local-dir ./hf_rlcr4reccon

# Download a specific trained model checkpoint (e.g. RLCR-CRED-8B-opt or RLCR-CRED-32B-opt)
hf download Alexps9yyy/rlcr4reccon --include "models/RLCR-CRED-8B-opt/*" --local-dir .

3. OOD IFQA End-to-End Benchmark Summary (eval.mix = 0.05, 10,000 Paired Bootstrap)

System Pipeline (Stage 1 + Stage 3 + Stage 5) Stage 2–5 Audit Rate Divergent p_rw AUROC Full reccon AUROC Full Acc ($\uparrow$) Full Brier ($\downarrow$)
Native 3-Chain Mean (35B+20B S1, no audit) 30.7% (387/1262) 69.90 92.62 84.47% 11.23
All-Native ReCCon (35B/20B S1 + 35B S3 + 35B S5) 30.7% (387/1262) 79.08 93.72 86.13% 10.38
Only-Stage-5 RLCR-CRED-8B-opt (s002 / s004) 30.7% (387/1262) 78.70 (Chain 73.99) 93.44 85.58% 10.46
Full-8B Structure-RLCR (Joint-8B + QA-8B + CRED-8B) 19.6% (249/1273) 79.36 ($p=0.0053$) 92.81 85.78% 10.74
Full-32B Structure-RLCR (s034) 15.6% (198/1273) 81.84 ($p=0.0229$) 94.09 87.43% 9.81
Full-32B Structure-RLCR (s038) 15.6% (198/1273) 81.13 ($p=0.0471$) 94.17 87.04% 9.93
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for Alexps9yyy/rlcr4reccon

Base model

Qwen/Qwen3-32B
Finetuned
(608)
this model