Unify Florence and Qwen Bridge Rebuttal seeds
Browse files- AGENTS.md +1 -0
- BRIDGE_SEED_PROTOCOL.md +28 -0
- EXPERIMENT_CONTRACT.md +3 -1
- README.md +13 -3
- REBUTTAL_EXPERIMENT_PLAN.md +2 -0
- qwen_bridge_common_start_29k/AGENTS.md +4 -5
- qwen_bridge_common_start_29k/MULTISEED_PROTOCOL.md +5 -69
- qwen_bridge_common_start_29k/README.md +4 -9
- qwen_bridge_common_start_29k/checkpoint_metadata_no_sha.json +6 -1
AGENTS.md
CHANGED
|
@@ -59,6 +59,7 @@ Use `scripts/evaluate_bridge_384_no_video.sh`. The locked panel has policy seeds
|
|
| 59 |
|
| 60 |
## Where to look first
|
| 61 |
|
|
|
|
| 62 |
- Contract and factual provenance: `EXPERIMENT_CONTRACT.md`
|
| 63 |
- File roles and sizes: `ASSET_MANIFEST.json`
|
| 64 |
- Source authority and known corrections: `SOURCE_INVENTORY.md`
|
|
|
|
| 59 |
|
| 60 |
## Where to look first
|
| 61 |
|
| 62 |
+
- Bridge training and evaluation seeds for Rebuttal: `BRIDGE_SEED_PROTOCOL.md`
|
| 63 |
- Contract and factual provenance: `EXPERIMENT_CONTRACT.md`
|
| 64 |
- File roles and sizes: `ASSET_MANIFEST.json`
|
| 65 |
- Source authority and known corrections: `SOURCE_INVENTORY.md`
|
BRIDGE_SEED_PROTOCOL.md
ADDED
|
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# ReTime Rebuttal:Bridge 训练与测评种子
|
| 2 |
+
|
| 3 |
+
## Florence–Bridge
|
| 4 |
+
|
| 5 |
+
20K common start 的基础训练 seed 为 `42`。从该 checkpoint 分支训练 500 步,主实验的 continuation/sampler seed 为 `117465`。已报告的多训练种子实验使用以下三组,所有组均从同一 20K 起点分支。
|
| 6 |
+
|
| 7 |
+
| 已报告的 continuation / sampler seed | Model RNG seed | Transform RNG seed |
|
| 8 |
+
|---|---:|---:|
|
| 9 |
+
| `117465`(主实验) | `588643` | `298226` |
|
| 10 |
+
| `533389` | `660882` | `812639` |
|
| 11 |
+
| `944964` | `659154` | `584920` |
|
| 12 |
+
|
| 13 |
+
三个 continuation seed 共用正式测评 policy seeds **`518957, 765353, 686765, 516408`**。每个 checkpoint 在 Stack、Carrot、Spoon、Eggplant 上各测 24 episodes / policy seed,共 384 rollouts,BF16、无视频。
|
| 14 |
+
|
| 15 |
+
Rebuttal 的 Florence 单种子复现沿用主实验的 continuation seed `117465`、model RNG `588643`、transform RNG `298226` 及上述四个 policy seeds。多训练种子汇总以表中三组为准。
|
| 16 |
+
|
| 17 |
+
训练种子取自实际 run 配置,测评口径以当时记录的 [Florence–Bridge 多训练种子结果 CSV](https://github.com/miaochengy-pngu/OPTR-StarVLA/blob/main/experiments/results/csv/florence_bridge_multiple_training_seeds_eval_results.csv) 为准。
|
| 18 |
+
|
| 19 |
+
## Qwen3-VL-4B–Bridge
|
| 20 |
+
|
| 21 |
+
29K common start 的基础训练 seed 为 `42`。历史 29K → 29.5K continuation 配置保留 `seed: 42`,从 29K 训练状态继承 RNG 和数据抽取位置。Rebuttal 若追加主动重新设种子的 Qwen continuation,具体训练 seed 清单待定,同一组 seed 用于该组所有配对运行。
|
| 22 |
+
|
| 23 |
+
| 要复现的记录 | Evaluation policy seeds | 每 checkpoint |
|
| 24 |
+
|---|---|---:|
|
| 25 |
+
| 29K common-start 评测 | `518957, 765353, 686765, 516408` | 384 |
|
| 26 |
+
| 29.5K OPTR 已报告评测 | `641297, 765353, 869156, 441297` | 384 |
|
| 27 |
+
|
| 28 |
+
两套均为四个 policy seeds × 四个任务 × 24 episodes。测评种子见 [Bridge 主结果 CSV](https://github.com/miaochengy-pngu/OPTR-StarVLA/blob/main/experiments/results/csv/main_bridge_results.csv),基础训练种子见 [29K common-start 配置](https://huggingface.co/mmmmmcyaaa/ReTime/blob/main/qwen_bridge_common_start_29k/config.full.yaml)。
|
EXPERIMENT_CONTRACT.md
CHANGED
|
@@ -13,7 +13,7 @@ Matched-Random/shuffle is intentionally excluded from the uploaded checkpoint se
|
|
| 13 |
|
| 14 |
The parent is the Florence–Bridge 20K complete Accelerate checkpoint produced by the seed-42 scratch run. It contains the consolidated model, DeepSpeed ZeRO-2 optimizer shards, model-state shard, per-rank RNG states, `latest`, `checkpoint.json`, and `zero_to_fp32.py`.
|
| 15 |
|
| 16 |
-
The continuation restores model, optimizer, scheduler position, RNG streams
|
| 17 |
|
| 18 |
The published 20.5K Uniform and Residual files are model-only evaluation snapshots. They are not substitutes for the full parent state.
|
| 19 |
|
|
@@ -39,6 +39,8 @@ The published 20.5K Uniform and Residual files are model-only evaluation snapsho
|
|
| 39 |
| Seeds | parent 42; continuation/sampler 117465; model 588643; transform 298226 |
|
| 40 |
| Logging | every 10 steps in the portable handoff; historical main run logged every 50 |
|
| 41 |
|
|
|
|
|
|
|
| 42 |
No explicit `freeze_modules` or `trainable_modules` filter was configured. The optimizer contained the VLM-interface and action-model groups. In the actually executed action path, Florence vision/text interface parameters and the action head changed; decoder parameters outside that path received no update. Do not summarize this as “the entire VLM was frozen” or “every registered parameter was actively trained.”
|
| 43 |
|
| 44 |
## 4. Data and proposal
|
|
|
|
| 13 |
|
| 14 |
The parent is the Florence–Bridge 20K complete Accelerate checkpoint produced by the seed-42 scratch run. It contains the consolidated model, DeepSpeed ZeRO-2 optimizer shards, model-state shard, per-rank RNG states, `latest`, `checkpoint.json`, and `zero_to_fp32.py`.
|
| 15 |
|
| 16 |
+
The continuation restores the model, optimizer, scheduler position, data-order cursor and update counter from the parent. The external branch initializes its RNG streams with the configured continuation, model and transform seeds. It starts at completed step 20,000 and performs exactly 500 optimizer updates. The active learning rates at resume were `9.407412610098572e-06` for the VLM interface and `9.407412610098571e-05` for the action model. The restored scheduler was `LambdaLR(last_epoch=20000, step_count=20001)`.
|
| 17 |
|
| 18 |
The published 20.5K Uniform and Residual files are model-only evaluation snapshots. They are not substitutes for the full parent state.
|
| 19 |
|
|
|
|
| 39 |
| Seeds | parent 42; continuation/sampler 117465; model 588643; transform 298226 |
|
| 40 |
| Logging | every 10 steps in the portable handoff; historical main run logged every 50 |
|
| 41 |
|
| 42 |
+
See the [Bridge seed protocol](BRIDGE_SEED_PROTOCOL.md) for the reported multi-seed panel and the seeds used in Rebuttal reproductions.
|
| 43 |
+
|
| 44 |
No explicit `freeze_modules` or `trainable_modules` filter was configured. The optimizer contained the VLM-interface and action-model groups. In the actually executed action path, Florence vision/text interface parameters and the action head changed; decoder parameters outside that path received no update. Do not summarize this as “the entire VLM was frozen” or “every registered parameter was actively trained.”
|
| 45 |
|
| 46 |
## 4. Data and proposal
|
README.md
CHANGED
|
@@ -14,6 +14,16 @@ tags:
|
|
| 14 |
|
| 15 |
This package contains the server-authoritative Florence–Bridge OPTR code, its complete 20K common-start state and Uniform/Residual-OPTR 20.5K evaluation snapshots. It also contains the complete Qwen3-VL-4B–Bridge 29K common-start state and the environment/seed contract needed for paired multi-seed continuation experiments.
|
| 16 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
## What is included
|
| 18 |
|
| 19 |
- `common_start_20k/checkpoints/steps_20000_accelerate/`: **full resumable state** (model, optimizer, scheduler metadata and per-rank RNG state). Start every new continuation arm here.
|
|
@@ -33,11 +43,11 @@ This package contains the server-authoritative Florence–Bridge OPTR code, its
|
|
| 33 |
|
| 34 |
The 20.5K files are sufficient for evaluation but **not** for exact optimizer/RNG continuation. Any new branch must resume from the complete 20K state.
|
| 35 |
|
| 36 |
-
For Qwen–Bridge, exact continuation requires **both** 29K `.pt` files.
|
| 37 |
|
| 38 |
This release does not use implementation, code-path, absolute-path, current-manifest or self-referential SHA gates. Experimental identity is enforced through loadability and semantic model/data/episode/anchor/support/batch/budget checks.
|
| 39 |
|
| 40 |
-
##
|
| 41 |
|
| 42 |
Fixed no-video SimplerEnv WidowX/Bridge panel: four policy seeds (`518957`, `765353`, `686765`, `516408`) × four tasks × 24 episodes = 384 rollouts.
|
| 43 |
|
|
@@ -46,7 +56,7 @@ Fixed no-video SimplerEnv WidowX/Bridge panel: four policy seeds (`518957`, `765
|
|
| 46 |
| Uniform +500 | 13/96 | 31/96 | 50/96 | 68/96 | 162/384 = 42.19% |
|
| 47 |
| Residual-OPTR +500 | 24/96 | 45/96 | 83/96 | 76/96 | 228/384 = 59.38% |
|
| 48 |
|
| 49 |
-
##
|
| 50 |
|
| 51 |
1. Download this repository and install `requirements.txt` in the same software environment used for StarVLA.
|
| 52 |
2. Download Florence-2-large and the OXE LeRobot Bridge/RT1 data.
|
|
|
|
| 14 |
|
| 15 |
This package contains the server-authoritative Florence–Bridge OPTR code, its complete 20K common-start state and Uniform/Residual-OPTR 20.5K evaluation snapshots. It also contains the complete Qwen3-VL-4B–Bridge 29K common-start state and the environment/seed contract needed for paired multi-seed continuation experiments.
|
| 16 |
|
| 17 |
+
## Rebuttal 训练与测评种子
|
| 18 |
+
|
| 19 |
+
| 设置 | Common-start 训练 seed | Continuation 训练 seed | Evaluation policy seeds |
|
| 20 |
+
|---|---|---|---|
|
| 21 |
+
| Florence–Bridge 20K → 20.5K | `42` | 主实验 `117465`,已报告三组 `117465, 533389, 944964` | `518957, 765353, 686765, 516408` |
|
| 22 |
+
| Qwen3-VL-4B–Bridge 29K common start | `42` | 历史 +500 配置为 `42`,继承 29K RNG。新多种子清单待定 | `518957, 765353, 686765, 516408` |
|
| 23 |
+
| Qwen3-VL-4B–Bridge 29.5K OPTR 已报告评测 | `42` | 同上 | `641297, 765353, 869156, 441297` |
|
| 24 |
+
|
| 25 |
+
完整的 Florence RNG 字段、Rebuttal 复用规则和原始记录链接见 [Bridge 训练与测评种子总表](BRIDGE_SEED_PROTOCOL.md)。
|
| 26 |
+
|
| 27 |
## What is included
|
| 28 |
|
| 29 |
- `common_start_20k/checkpoints/steps_20000_accelerate/`: **full resumable state** (model, optimizer, scheduler metadata and per-rank RNG state). Start every new continuation arm here.
|
|
|
|
| 43 |
|
| 44 |
The 20.5K files are sufficient for evaluation but **not** for exact optimizer/RNG continuation. Any new branch must resume from the complete 20K state.
|
| 45 |
|
| 46 |
+
For Qwen–Bridge, exact continuation requires **both** 29K `.pt` files.
|
| 47 |
|
| 48 |
This release does not use implementation, code-path, absolute-path, current-manifest or self-referential SHA gates. Experimental identity is enforced through loadability and semantic model/data/episode/anchor/support/batch/budget checks.
|
| 49 |
|
| 50 |
+
## Florence–Bridge headline result
|
| 51 |
|
| 52 |
Fixed no-video SimplerEnv WidowX/Bridge panel: four policy seeds (`518957`, `765353`, `686765`, `516408`) × four tasks × 24 episodes = 384 rollouts.
|
| 53 |
|
|
|
|
| 56 |
| Uniform +500 | 13/96 | 31/96 | 50/96 | 68/96 | 162/384 = 42.19% |
|
| 57 |
| Residual-OPTR +500 | 24/96 | 45/96 | 83/96 | 76/96 | 228/384 = 59.38% |
|
| 58 |
|
| 59 |
+
## Florence–Bridge quick start
|
| 60 |
|
| 61 |
1. Download this repository and install `requirements.txt` in the same software environment used for StarVLA.
|
| 62 |
2. Download Florence-2-large and the OXE LeRobot Bridge/RT1 data.
|
REBUTTAL_EXPERIMENT_PLAN.md
CHANGED
|
@@ -2,6 +2,8 @@
|
|
| 2 |
|
| 3 |
本文档记录与师兄 meeting 后确定的补充实验方向。后续执行时以服务器上的真实代码、配置、完整 checkpoint 和日志为准;所有方法比较必须共享数据、训练预算、训练种子和评测协议,除目标变量外不得改变其他条件。
|
| 4 |
|
|
|
|
|
|
|
| 5 |
## 1. 周期性刷新 proposal 的成本—性能曲线
|
| 6 |
|
| 7 |
从同一个完整 common-start checkpoint 出发,每继续训练 500 steps 后重新计算并刷新一次 residual proposal,再继续训练下一段 500 steps。重复多个阶段,并在每个阶段保存 checkpoint、执行固定协议评测。
|
|
|
|
| 2 |
|
| 3 |
本文档记录与师兄 meeting 后确定的补充实验方向。后续执行时以服务器上的真实代码、配置、完整 checkpoint 和日志为准;所有方法比较必须共享数据、训练预算、训练种子和评测协议,除目标变量外不得改变其他条件。
|
| 4 |
|
| 5 |
+
Florence–Bridge 和 Qwen–Bridge 的训练及测评种子统一见 [Bridge 训练与测评种子总表](BRIDGE_SEED_PROTOCOL.md)。
|
| 6 |
+
|
| 7 |
## 1. 周期性刷新 proposal 的成本—性能曲线
|
| 8 |
|
| 9 |
从同一个完整 common-start checkpoint 出发,每继续训练 500 steps 后重新计算并刷新一次 residual proposal,再继续训练下一段 500 steps。重复多个阶段,并在每个阶段保存 checkpoint、执行固定协议评测。
|
qwen_bridge_common_start_29k/AGENTS.md
CHANGED
|
@@ -1,24 +1,23 @@
|
|
| 1 |
# Qwen–Bridge 29K Common Start 操作约束
|
| 2 |
|
| 3 |
-
本目录
|
| 4 |
|
| 5 |
## 不可改变
|
| 6 |
|
| 7 |
- 起点必须同时加载 `steps_29000_pytorch_model.pt` 与 `steps_29000_training_state.pt`。
|
| 8 |
- common-start 历史训练 seed 是 `42`。
|
| 9 |
- 每个主动 reseed realization 下,各比较方法必须共享同一个 continuation training seed。
|
| 10 |
-
- paired continuation
|
| 11 |
-
-
|
| 12 |
- 每个 checkpoint 的正式测评是 4 seeds × 4 tasks × 24 episodes = 384 rollouts。
|
| 13 |
- global batch、学习率、scheduler、训练预算、world size、数据与 support 合同必须在方法间保持一致。
|
| 14 |
|
| 15 |
## 禁止
|
| 16 |
|
| 17 |
- 禁止把 evaluation seed 当成 training seed。
|
| 18 |
-
-
|
| 19 |
- 禁止只恢复模型而丢弃 optimizer、scheduler、RNG 或 data-order。
|
| 20 |
- 禁止新增代码、实现、绝对路径、manifest 自指等工程哈希启动门。
|
| 21 |
-
- 禁止把计划中的四个 training seeds 写成已经完成的 Qwen 历史结果。
|
| 22 |
|
| 23 |
## 允许的可移植性修改
|
| 24 |
|
|
|
|
| 1 |
# Qwen–Bridge 29K Common Start 操作约束
|
| 2 |
|
| 3 |
+
本目录用于从固定 29K common start 开展 continuation 实验。操作前阅读 `README.md`、`ENVIRONMENT.md`、`MULTISEED_PROTOCOL.md` 和 [Bridge 训练与测评种子总表](https://huggingface.co/mmmmmcyaaa/ReTime/blob/main/BRIDGE_SEED_PROTOCOL.md)。
|
| 4 |
|
| 5 |
## 不可改变
|
| 6 |
|
| 7 |
- 起点必须同时加载 `steps_29000_pytorch_model.pt` 与 `steps_29000_training_state.pt`。
|
| 8 |
- common-start 历史训练 seed 是 `42`。
|
| 9 |
- 每个主动 reseed realization 下,各比较方法必须共享同一个 continuation training seed。
|
| 10 |
+
- 新的 Qwen paired continuation seed 清单待定。
|
| 11 |
+
- 复现历史评测时,按总表选对应 checkpoint 的 policy seeds。同一配对评测内固定使用相同的 policy seeds。
|
| 12 |
- 每个 checkpoint 的正式测评是 4 seeds × 4 tasks × 24 episodes = 384 rollouts。
|
| 13 |
- global batch、学习率、scheduler、训练预算、world size、数据与 support 合同必须在方法间保持一致。
|
| 14 |
|
| 15 |
## 禁止
|
| 16 |
|
| 17 |
- 禁止把 evaluation seed 当成 training seed。
|
| 18 |
+
- 区分历史 `seed=42` 的完整恢复与未来主动 reseed。
|
| 19 |
- 禁止只恢复模型而丢弃 optimizer、scheduler、RNG 或 data-order。
|
| 20 |
- 禁止新增代码、实现、绝对路径、manifest 自指等工程哈希启动门。
|
|
|
|
| 21 |
|
| 22 |
## 允许的可移植性修改
|
| 23 |
|
qwen_bridge_common_start_29k/MULTISEED_PROTOCOL.md
CHANGED
|
@@ -1,73 +1,9 @@
|
|
| 1 |
-
# Qwen–Bridge
|
| 2 |
|
| 3 |
-
|
| 4 |
|
| 5 |
-
|
| 6 |
|
| 7 |
-
|
| 8 |
-
- 作用:生成本交接包中的 29K common start。
|
| 9 |
-
- 该 seed 已经固化在 29K checkpoint 的历史中,不是后续多种子实验的横轴。
|
| 10 |
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
历史首轮 29K→29.5K 训练采用完整状态恢复,没有主动覆盖 checkpoint 内的 RNG/data-order,因此应描述为:
|
| 14 |
-
|
| 15 |
-
> Exact-state continuation from the 29K common-start checkpoint.
|
| 16 |
-
|
| 17 |
-
为回答多训练种子稳定性问题,计划使用以下四个 paired continuation seeds:
|
| 18 |
-
|
| 19 |
-
```text
|
| 20 |
-
117465
|
| 21 |
-
239217
|
| 22 |
-
533389
|
| 23 |
-
587008
|
| 24 |
-
```
|
| 25 |
-
|
| 26 |
-
这四个 seed 是 **计划执行的 reseeded continuation panel**,不能写成已经由本 Qwen 29K checkpoint 完成的历史结果。每个 seed 下,所有比较方法必须使用同一个 continuation seed;禁止给不同方法分配不同 seed。
|
| 27 |
-
|
| 28 |
-
主动 reseed 时必须统一覆盖:
|
| 29 |
-
|
| 30 |
-
- Python random;
|
| 31 |
-
- NumPy;
|
| 32 |
-
- PyTorch CPU;
|
| 33 |
-
- PyTorch CUDA;
|
| 34 |
-
- dataloader/data-order;
|
| 35 |
-
- source 与 trajectory sampling;
|
| 36 |
-
- temporal-anchor sampling;
|
| 37 |
-
- Flow-Matching noise;
|
| 38 |
-
- dropout 和其他训练随机过程。
|
| 39 |
-
|
| 40 |
-
除 seed/RNG 外,模型、optimizer、scheduler、起始 step、数据、batch、训练预算、world size、学习率和保存协议保持不变。
|
| 41 |
-
|
| 42 |
-
### 1.3 Evaluation policy seed
|
| 43 |
-
|
| 44 |
-
正式 headline 测评固定:
|
| 45 |
-
|
| 46 |
-
```text
|
| 47 |
-
518957
|
| 48 |
-
765353
|
| 49 |
-
686765
|
| 50 |
-
516408
|
| 51 |
-
```
|
| 52 |
-
|
| 53 |
-
对每个 checkpoint:
|
| 54 |
-
|
| 55 |
-
- 4 个 policy seeds;
|
| 56 |
-
- Stack、Carrot、Spoon、Eggplant 四个任务;
|
| 57 |
-
- 每个 seed/task 24 episodes;
|
| 58 |
-
- 合计 384 rollouts;
|
| 59 |
-
- BF16 inference;
|
| 60 |
-
- 无视频。
|
| 61 |
-
|
| 62 |
-
这些 evaluation policy seeds 只控制测评 realization,不用于训练,也不能替代 continuation training seeds。
|
| 63 |
-
|
| 64 |
-
## 2. 推荐结果表
|
| 65 |
-
|
| 66 |
-
| Continuation seed | Uniform | Shuffle/Random | OPTR | 其他 baseline |
|
| 67 |
-
|---|---:|---:|---:|---:|
|
| 68 |
-
| 117465 | 待运行 | 待运行 | 待运行 | 待运行 |
|
| 69 |
-
| 239217 | 待运行 | 待运行 | 待运行 | 待运行 |
|
| 70 |
-
| 533389 | 待运行 | 待运行 | 待运行 | 待运行 |
|
| 71 |
-
| 587008 | 待运行 | 待运行 | 待运行 | 待运行 |
|
| 72 |
-
|
| 73 |
-
每个单元格都应来自同一套 384-rollout formal evaluation。最终报告逐 seed、均值、样本标准差、范围,以及配对差值 OPTR−Uniform、OPTR−Shuffle。
|
|
|
|
| 1 |
+
# Qwen–Bridge continuation 与测评种子
|
| 2 |
|
| 3 |
+
Florence 和 Qwen 的已报告训练、测评种子统一登记在 [Bridge 训练与测评种子总表](https://huggingface.co/mmmmmcyaaa/ReTime/blob/main/BRIDGE_SEED_PROTOCOL.md)。
|
| 4 |
|
| 5 |
+
29K common start 的基础训练 seed 为 `42`。历史 29K → 29.5K continuation 配置为 `seed: 42`,继承 checkpoint 中保存的 RNG 与数据抽取位置。
|
| 6 |
|
| 7 |
+
Rebuttal 若新增主动 reseed 的 Qwen continuation,训练 seed 清单待定。同一组的配对运行共享 continuation seed,并以相同方式初始化 Python、NumPy、PyTorch CPU/CUDA、dataloader、采样和模型 RNG。模型、optimizer、scheduler、起始 step、batch、学习率和训练预算沿用共同协议。
|
|
|
|
|
|
|
| 8 |
|
| 9 |
+
复现历史测评时,按总表选择对应 checkpoint 的 policy seed。每组评测为四个 policy seeds × Stack、Carrot、Spoon、Eggplant 四任务 × 24 episodes,共 384 rollouts,BF16、无视频。
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
qwen_bridge_common_start_29k/README.md
CHANGED
|
@@ -36,19 +36,14 @@ qwen_bridge_common_start_29k/
|
|
| 36 |
|
| 37 |
## Seed 一览
|
| 38 |
|
| 39 |
-
|
| 40 |
-
|---|---|---|
|
| 41 |
-
| common-start 基础训练 seed | `42` | 生成本 29K checkpoint 的训练 seed |
|
| 42 |
-
| 历史 29K→29.5K exact-state continuation | 不主动重设 | 直接继承 29K training state 内的 RNG 与 data-order |
|
| 43 |
-
| 计划的 paired continuation training seeds | `117465, 239217, 533389, 587008` | 多种子回复实验;每个 seed 下所有方法必须共用同一个 seed |
|
| 44 |
-
| 正式 evaluation policy seeds | `518957, 765353, 686765, 516408` | 每个 checkpoint 固定 4 seeds × 4 tasks × 24 episodes = 384 rollouts |
|
| 45 |
|
| 46 |
-
|
| 47 |
|
| 48 |
## 重要边界
|
| 49 |
|
| 50 |
-
1. `seed=42` 是 common-start
|
| 51 |
-
2. 直接载入完整 training state
|
| 52 |
3. 同一个 continuation seed 下,各方法必须同步重设 Python、NumPy、PyTorch、CUDA、data-order、source/trajectory sampling、temporal anchor sampling、Flow-Matching noise 和 dropout 等随机状态。
|
| 53 |
4. 测评 policy seed 与训练 seed 相互独立,禁止混用。
|
| 54 |
5. 不通过代码路径或哈希判断实验身份;保留真正影响语义的模型、数据集、episode、anchor/support、proposal、batch 和预算检查。
|
|
|
|
| 36 |
|
| 37 |
## Seed 一览
|
| 38 |
|
| 39 |
+
29K 的基础训练 seed 为 `42`。历史 +500 continuation 配置保留 `seed: 42`,继承 29K checkpoint 的 RNG 与数据抽取位置。Rebuttal 中新设的 Qwen continuation seed 清单待定。
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 40 |
|
| 41 |
+
两套历史评测 seed 及 Florence 的对应记录统一见 [Bridge 训练与测评种子总表](https://huggingface.co/mmmmmcyaaa/ReTime/blob/main/BRIDGE_SEED_PROTOCOL.md)。Qwen 恢复与配对规则见 [MULTISEED_PROTOCOL.md](MULTISEED_PROTOCOL.md)。
|
| 42 |
|
| 43 |
## 重要边界
|
| 44 |
|
| 45 |
+
1. `seed=42` 是 common-start 与历史 +500 continuation 的配置值。
|
| 46 |
+
2. 直接载入完整 training state 并继承 RNG 属于 exact-state continuation。若主动设置新 seed,则称为 *independent reseeded continuation replication from a fixed common-start checkpoint*。
|
| 47 |
3. 同一个 continuation seed 下,各方法必须同步重设 Python、NumPy、PyTorch、CUDA、data-order、source/trajectory sampling、temporal anchor sampling、Flow-Matching noise 和 dropout 等随机状态。
|
| 48 |
4. 测评 policy seed 与训练 seed 相互独立,禁止混用。
|
| 49 |
5. 不通过代码路径或哈希判断实验身份;保留真正影响语义的模型、数据集、episode、anchor/support、proposal、batch 和预算检查。
|
qwen_bridge_common_start_29k/checkpoint_metadata_no_sha.json
CHANGED
|
@@ -20,7 +20,12 @@
|
|
| 20 |
}
|
| 21 |
],
|
| 22 |
"historical_continuation_mode": "exact-state; no active reseed",
|
| 23 |
-
"
|
|
|
|
|
|
|
| 24 |
"formal_evaluation_policy_seeds": [518957, 765353, 686765, 516408],
|
|
|
|
|
|
|
|
|
|
| 25 |
"sha_gate": false
|
| 26 |
}
|
|
|
|
| 20 |
}
|
| 21 |
],
|
| 22 |
"historical_continuation_mode": "exact-state; no active reseed",
|
| 23 |
+
"historical_continuation_config_seed": 42,
|
| 24 |
+
"planned_paired_continuation_training_seeds": [],
|
| 25 |
+
"planned_paired_continuation_training_seed_status": "pending",
|
| 26 |
"formal_evaluation_policy_seeds": [518957, 765353, 686765, 516408],
|
| 27 |
+
"formal_evaluation_panel_scope": "29K common-start evaluation",
|
| 28 |
+
"reported_optr_29500_evaluation_policy_seeds": [641297, 765353, 869156, 441297],
|
| 29 |
+
"seed_protocol": "BRIDGE_SEED_PROTOCOL.md",
|
| 30 |
"sha_gate": false
|
| 31 |
}
|