Phase-13 fixes: MV-base GSM8K collision, GSM8K reasoning max_gen_toks=16000, IFEval loose, +3 missing cells (eval_responses.parquet)

Phase-13 fixes: MV-base GSM8K collision, GSM8K reasoning max_gen_toks=16000, IFEval loose, +3 missing cells (eval_scores.parquet)

GSM8K 3-seed sampled responses + per-sample is_correct (8/11 tasks now per-sample graded)

GSM8K: 3-seed sampled (temp0.7, do_sample) mean@3 + per-seed + std; replaces greedy single-seed

Add per-sample is_correct for JEEBench (exact) + IFEval (strict); now 7 tasks graded per-sample

Add per-sample is_correct for 5 reasoning tasks (MATH500/AIME24/AIME25/AMC23/GPQADiamond); per-seed means reproduce eval_scores exactly

AIME24/25/AMC23: Aug-30 sampled 3-seed responses (run_idx 0/1/2); replaces Aug-2 4-trial set

AIME24/25/AMC23: Aug-30 sampled 3-seed (mean@3 + per-seed + std, num_repeat=3); replaces Aug-2 4-trial set

Adopt gsm8k_paths taxonomy: relabel Qwen2.5/Qwen3 Instruct-OT30K rows (28392 rows Base/OT3 -> Instruct/OT3)

Adopt gsm8k_paths taxonomy: relabel Qwen2.5/Qwen3 Instruct-OT30K (sft__ot30k_Qwen* Base/OT3 -> Instruct/OT3)

Add per-seed scores (score_seed0/1/2 + score_std + score_std_err) for the 7 seed3-folded tasks; mean stays in score

Update eval_scores to 3-seed mean@3: HumanEval/MBPP/IFEval x92 ckpts + GPQA/JEE/MATH500 straggler (num_repeat 1->3, 279 rows)

Fold seed3 mean@3 sampling: replace n=1 with 3-seed outputs for MATH500/GPQA/JEE/LCB/HumanEval/MBPP/IFEval (1,357,733 rows)

seed3: mean@3 sampling for MATH500/GPQADiamond/JEEBench/LiveCodeBenchv5 (369 cells; MATH500 re-graded over 3 seeds; 92/93 checkpoints)

Fold in 13 recent ablation checkpoints (MV-100B base/SFT/DPO, woOT3-100B/300B bases, Nemotron-100B plain+mvR base/SFT/OT3) — greedy+sampling math, code+gsm8k sampling. +221 score rows, +75k responses.

Fold in 13 recent ablation checkpoints (MV-100B base/SFT/DPO, woOT3-100B/300B bases, Nemotron-100B plain+mvR base/SFT/OT3) — greedy+sampling math, code+gsm8k sampling. +221 score rows, +75k responses.

Fold-in: OT3-family code sampling rows upgraded to 32k (truncation-free) for HumanEval/MBPP/IFEval/LCB; 76 cells (19 OT3/OT3+SFT/OT3+DPO ckpts). Non-OT3 stay 16k.

Fold-in: OT3-family code sampling rows upgraded to 32k (truncation-free) for HumanEval/MBPP/IFEval/LCB; 76 cells (19 OT3/OT3+SFT/OT3+DPO ckpts). Non-OT3 stay 16k.

Add decoding=sampling for HumanEval/MBPP/IFEval/LiveCodeBenchv5/GSM8K across 80 checkpoints (16k code / 8k gsm8k, T=0.7); +228k response rows, +400 score rows

Add decoding=sampling for HumanEval/MBPP/IFEval/LiveCodeBenchv5/GSM8K across 80 checkpoints (16k code / 8k gsm8k, T=0.7); +228k response rows, +400 score rows

Add dual-regime (greedy/sampling) evals @ max_tokens=32000 for 80 checkpoints (6 tasks); add decoding+max_tokens cols; legacy rows tagged decoding=legacy; +eval_scores_dual.csv

Add dual-regime (greedy/sampling) evals @ max_tokens=32000 for 80 checkpoints (6 tasks); add decoding+max_tokens cols; legacy rows tagged decoding=legacy; +eval_scores_dual.csv

Add dual-regime (greedy/sampling) evals @ max_tokens=32000 for 80 checkpoints (6 tasks); add decoding+max_tokens cols; legacy rows tagged decoding=legacy; +eval_scores_dual.csv

Scores: +MV-100B ablation, corrected code caps.

Scores: +MV-100B ablation, corrected code caps.

Add MV-100B OT-in-pretraining ablation (5 models, with/wo OT3 pretrain); corrected MBPP/IFEval @8192. 311951 rows.

Refresh scores: all 6 bases x SR/NC/AR + corrected greedy code.

Refresh scores: all 6 bases x SR/NC/AR + corrected greedy code.

Add 24 non-MV NC/AR reasoning models; corrected greedy MBPP/IFEval (512); all 6 bases x SR/NC/AR. 291126 rows.

Fill IFEval gaps (12 cells, 4096-tok re-runs) + NC+SFT LCB; 189666 rows, all 11 tasks complete, with model_answer_extracted.

Refresh scores: +IFEval re-runs (4096 tok) + NC+SFT LCB.

Refresh scores: +IFEval re-runs (4096 tok) + NC+SFT LCB.

Improve code extraction (full-text fence search + looser fallback): HumanEval 80->90%, MBPP 72%, LCB 48%.

Add reasoning-aware model_answer_extracted column (strip <think>, then boxed/number/code/letter per task). ~99% math/MCQ, 80% HE, keeps raw model_answer.

Add eval scores (all bases x stages x 11 tasks): long parquet + wide CSV.

Add eval scores (all bases x stages x 11 tasks): long parquet + wide CSV.

Add temp-0.7 x3 AIME24+AIME25 responses (6 MV stages) alongside existing; new `decoding` column distinguishes (sample_temp0.7 vs original).

Add all-11-task responses: SR/NC/AR reasoning stages (math+gsm8k) + code/IFEval (HumanEval/MBPP/IFEval/LCB) backfill for all models. 69107->182264 rows.

Upload mv_aime24_amc23_best_responses.parquet with huggingface_hub

Upload eval responses (7 tasks, 11 bases, 6 stages, 69k rows)

initial commit