# Alignment with paper 338

The supplied paper's Table 1 (page 6) defines the main system as heterogeneous Qwen3-1.7B / Qwen3-4B / Qwen3-8B with independent rank-64 LoRA adapters. The homogeneous 3x4B and 3x8B experiments, baselines and ablations are not part of this release. Tables 15–17 (pages 32–34) identify the final PD and SLS data counts and optimization settings. The following mapping was checked against the source project's `final_hyperparameter_audit.md`, its extractor and checkpoint metadata.

## Paper-reported main results

| Dataset | Metric 1 | Metric 2 | Evaluation examples |
|---|---:|---:|---:|
| MuSiQue | Accuracy 42.37 | F1 53.43 | 2,417 |
| GSM-Hard | Accuracy 71.97 | F1 81.15 | 132 |
| MATH | Accuracy 78.40 | F1 85.25 | 500 |
| MultiPL-E | Weighted pass@1 59.39 | Macro-language pass@1 59.63 | 1,352 |
| Conifer | Coverage 88.90 | Explicit 95.90 | 1,402 |

All scores are percentages reported in the paper. They are not measurements newly produced during publication. In particular, the MultiPL-E 8B evaluated checkpoint is missing, so the included later weights do not substantiate that exact result.

## Final training-run mapping

Paths below are relative to the original `jca` project. Each listed model-stage group includes every surviving checkpoint weight under that final training version.

| Dataset / stage | Source run | Models / training agents |
|---|---|---|
| MuSiQue PD | `sft_runs/0706v2` | 1.7B A1, 4B A2, 8B A3 |
| MuSiQue SLS | `rl_runs/rl_14B_0716_v2` | 1.7B A1, 4B A2 |
| MuSiQue SLS | `rl_runs/rl_14B_0716_v3` | 8B A3 |
| GSM-Hard PD | `sft_runs/gsm_fixed_a1_corr30_20260723_v3` | 1.7B A1, 4B A2, 8B A3 |
| GSM-Hard SLS | `rl_runs/gsm_judge_rl/gsm_judge_rl_v13_14b_role_c2c_1_05_1` | 1.7B A1, 4B A2, 8B A3 |
| MATH PD | `experiments/math_specific_sft_rl_v1/artifacts/math_specific_sft_rl_reuse_legacy_v4d_real_rollout_sft_protocol_floor_v2_20260830/01_sft/adapters` | 1.7B A1, 4B A2, 8B A3 |
| MATH SLS | `experiments/math_specific_sft_rl_v1/artifacts/math_specific_sft_rl_reuse_legacy_v4d_real_rollout_sft_protocol_floor_v2_20260830/05_train/adapters` | 1.7B A1, 4B A2, 8B A3 |
| MultiPL-E PD | `sft_runs/multipl_e_8lang_v5_diverse_a1_5ep` | 1.7B A1 |
| MultiPL-E PD | `sft_runs/multipl_e_8lang_v5_diverse_a2_selfcorr2x_2ep_20260819` | 4B A2 |
| MultiPL-E PD | `sft_runs/multipl_e_8lang_v5_diverse_a3_5ep` | 8B A3 |
| MultiPL-E SLS | `rl_runs/multipl_e_a2_gpt5_paired_rwr_20260827` | 4B A2 |
| MultiPL-E SLS | `rl_runs/multipl_e_a1_8b_gpt5_paired_rwr_20260830` | 8B A1 |
| Conifer PD | `conifer_training_hub/11_runs/sft/conifer_stratified_sft_full_rl_20260902` | 1.7B A1, 4B A2, 8B A3 |
| Conifer SLS | `conifer_training_hub/11_runs/rl/conifer_stratified_sft_full_rl_20260902` | 1.7B A1, 4B A2, 8B A3 |

This produces 29 model-stage groups: 15 PD groups and 14 SLS groups. MultiPL-E 1.7B is PD-only, exactly as Table 16 specifies.

## Table 16 cross-check

The recorded final SLS summaries agree with these paper training-record counts:

| Dataset | 1.7B | 4B | 8B |
|---|---:|---:|---:|
| MuSiQue | 2,416 | 1,727 | 1,411 |
| GSM-Hard | 3,714 | 3,516 | 2,765 |
| MATH | 32,746 | 29,261 | 30,200 |
| MultiPL-E | PD only | 1,908 | 1,860 |
| Conifer | 46,746 | 47,063 | 44,913 |

## Evaluation checkpoint identity

- MuSiQue: A1 and A2 use `rl_14B_0716_v2`; A3 uses `rl_14B_0716_v3`. The formal evaluation identifier is `rl_sft_mas_0717_1130`.
- GSM-Hard: the formal evaluation uses the `gsm_judge_rl_v13` role-c2c run with seed 43, hidden thinking and the recorded self-handoff/context configuration.
- MATH: the formal evaluation identifier is `rl_all_final_a3_turn2_incumbent_v1`. Its published score uses the project's special three-turn answer-selection protocol; loading an adapter by itself does not reproduce that protocol.
- MultiPL-E: evaluation roles are reversed in capacity: A1=8B, A2=4B, A3=1.7B. The formal run used 8B SLS `checkpoint-25`, 4B SLS `checkpoint-120`, and 1.7B PD `final`. The 8B checkpoint 25 is missing; its later checkpoint 100 and final export are retained only as other checkpoints of the same final training run. The formal evaluation identifier is `mpe0912_120_multipl_e_gpt5rwr_ckpt25_x10_r08`.
- Conifer: the formal evaluation identifier is `rl_mas_eval_forever_20260906/round_4` from the final stratified SFT/RL run. The Coverage/Explicit scores must not be substituted with other per-trajectory hard-score or all-explicit-pass metrics.

## Known missing historical checkpoints

The source cleanup log records 18 deleted checkpoints from these selected runs on 2026-11-05, including MultiPL-E 8B checkpoint 25. Their exact paths and deletion timestamps are included in `manifest.json`. No substitute is labeled as the deleted checkpoint. Other earlier checkpoints can be absent because training used checkpoint-retention limits.