Add files using upload-large-folder tool

.pytest_cache/README.md

@@ -0,0 +1,8 @@
+# pytest cache directory #
+
+This directory contains data from the pytest's cache plugin,
+which provides the `--lf` and `--ff` options, as well as the `cache` fixture.
+
+**Do not** commit this to version control.
+
+See [the docs](https://docs.pytest.org/en/stable/how-to/cache.html) for more information.

results/ppl_rescore_20260923/pythia-1b/lm1b/mdlm_sdtt/steps_1.json

@@ -0,0 +1,31 @@
+{
+  "dataset": "lm1b",
+  "evaluation": "independent_unconditional_ppl_rescore",
+  "generative_perplexity": {
+    "aggregation": "sum_nll_over_sum_valid_next_tokens",
+    "batch_size": 8,
+    "context_length": 1024,
+    "forward_dtype": "float32",
+    "loss_dtype": "float32",
+    "model_id": "EleutherAI/pythia-1b",
+    "model_revision": "f73d7dcc545c8bd326d8559c8ef84ffe92fea6b2",
+    "model_snapshot": "/home/ruizzhao/.cache/huggingface/hub/models--EleutherAI--pythia-1b/snapshots/f73d7dcc545c8bd326d8559c8ef84ffe92fea6b2",
+    "offline_scoring": true,
+    "perplexity": 919.9931469085786,
+    "sample_count": 1024,
+    "tokenizer_id": "EleutherAI/pythia-1b",
+    "tokenizer_revision": "f73d7dcc545c8bd326d8559c8ef84ffe92fea6b2",
+    "tokenizer_snapshot": "/home/ruizzhao/.cache/huggingface/hub/models--EleutherAI--pythia-1b/snapshots/f73d7dcc545c8bd326d8559c8ef84ffe92fea6b2",
+    "torch_version": "2.8.0+cu128",
+    "total_nll": 987205.9931640625,
+    "transformers_version": "4.57.1",
+    "truncation": "right",
+    "valid_token_count": 144659
+  },
+  "generator": "mdlm_sdtt",
+  "sample_count": 1024,
+  "samples": "/home/ruizzhao/diffusion_baseline/results/samples/lm1b/mdlm_sdtt/steps_1/samples.jsonl",
+  "samples_sha256": "67edbe5a58df7c44186d924dd2d50a8e78397b546c1412fe8bf9efe6a7c246ab",
+  "schema_version": 1,
+  "steps": 1
+}

results/smoke/samples/owt/duo/steps_1/upstream_hydra/.hydra/config.yaml

@@ -0,0 +1,151 @@
+mode: sample_eval
+seed: 42
+loader:
+  global_batch_size: 512
+  eval_global_batch_size: ${.global_batch_size}
+  batch_size: 2
+  eval_batch_size: 8
+  num_workers: ${eval:"len(__import__('os').sched_getaffinity(0))"}
+  pin_memory: true
+sampling:
+  predictor: ancestral
+  steps: 1
+  noise_removal: ancestral
+  use_float64: true
+  p_nucleus: 1.0
+  num_sample_batches: 1
+  num_sample_log: 2
+  semi_ar: false
+  stride_length: 1
+  num_strides: 1
+  guid_weight: null
+  psi:
+    time_profile: linear
+    high_mode: pure-posterior
+    middle_mode: pure-posterior
+    low_mode: pure-posterior
+    high_frac: 0.2
+    middle_frac: 0.6
+training:
+  ema: 0.9999
+  antithetic_sampling: true
+  importance_sampling: false
+  sampling_eps: 0.001
+  change_of_variables: false
+  loss_precision: bf16
+  finetune_path: ''
+  class_dropout_p: 0.1
+eval:
+  checkpoint_path: /home/ruizzhao/diffusion_baseline/checkpoints/official/duo/owt
+  disable_ema: false
+  compute_generative_perplexity: false
+  perplexity_batch_size: 8
+  compute_perplexity_on_sanity: false
+  gen_ppl_eval_model_name_or_path: gpt2-large
+  generate_samples: true
+  generated_samples_path: /home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/duo/steps_1/upstream_samples.json
+optim:
+  weight_decay: 0
+  lr: 0.0003
+  beta1: 0.9
+  beta2: 0.999
+  eps: 1.0e-08
+trainer:
+  _target_: lightning.Trainer
+  accelerator: cuda
+  num_nodes: 1
+  devices: 1
+  accumulate_grad_batches: ${div_up:${loader.global_batch_size}, ${eval:${trainer.devices}
+    * ${loader.batch_size} * ${trainer.num_nodes}}}
+  gradient_clip_val: 1.0
+  precision: bf16
+  num_sanity_val_steps: 2
+  max_steps: 1000000
+  log_every_n_steps: 100
+  limit_train_batches: 1.0
+  limit_val_batches: 1.0
+  val_check_interval: 5000
+  check_val_every_n_epoch: null
+wandb:
+  project: duo
+  group: null
+  job_type: null
+  name: null
+  id: ${.name}_${seed}
+  tags:
+  - ${noise.type}
+  - ${data.train}
+  - ${data.valid}
+  - ${algo.name}
+  offline: true
+checkpointing:
+  save_dir: /home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/duo/steps_1/upstream_hydra
+  resume_from_ckpt: true
+  resume_ckpt_path: ${.save_dir}/checkpoints/last.ckpt
+callbacks:
+  checkpoint_every_n_steps:
+    _target_: lightning.pytorch.callbacks.ModelCheckpoint
+    save_top_k: -1
+    save_last: true
+    dirpath: ${checkpointing.save_dir}/checkpoints
+    verbose: true
+    auto_insert_metric_name: false
+    every_n_train_steps: 500
+  checkpoint_monitor:
+    _target_: lightning.pytorch.callbacks.ModelCheckpoint
+    monitor: val/nll
+    mode: min
+    save_top_k: 1
+    save_last: false
+    dirpath: ${checkpointing.save_dir}/checkpoints
+    filename: best
+    auto_insert_metric_name: false
+    verbose: true
+  learning_rate_monitor:
+    _target_: lightning.pytorch.callbacks.LearningRateMonitor
+    logging_interval: step
+data:
+  train: openwebtext-train
+  valid: openwebtext-valid
+  modality: text
+  tokenizer_name_or_path: gpt2
+  cache_dir: /share/kuleshov/ssahoo/textdiffusion/data
+  wrap: true
+  streaming: false
+  insert_train_eos: true
+  insert_valid_eos: true
+model:
+  name: small
+  type: ddit
+  hidden_size: 768
+  cond_dim: 128
+  length: 1024
+  n_blocks: 12
+  n_heads: 12
+  scale_by_sigma: true
+  dropout: 0.1
+  tie_word_embeddings: false
+  vocab_lookup: true
+strategy:
+  _target_: lightning.pytorch.strategies.DDPStrategy
+  find_unused_parameters: false
+noise:
+  type: log-linear
+  eps: 0.001
+lr_scheduler:
+  _target_: transformers.get_constant_schedule_with_warmup
+  num_warmup_steps: 2500
+prior:
+  type: none
+  latent_width: 0
+  latent_height: 0
+algo:
+  name: duo_base
+  backbone: hf_dit
+  parameterization: mean
+  time_conditioning: true
+  T: 0
+  subs_masking: false
+  causal_attention: false
+  ignore_bos: false
+  loss_type: elbo

results/smoke/samples/owt/duo/steps_1/upstream_hydra/.hydra/hydra.yaml

@@ -0,0 +1,181 @@
+hydra:
+  run:
+    dir: /home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/duo/steps_1/upstream_hydra
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.run.dir=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/duo/steps_1/upstream_hydra
+    - hydra.mode=RUN
+    task:
+    - mode=sample_eval
+    - seed=42
+    - data=openwebtext-split
+    - model=small
+    - model.length=1024
+    - loader.batch_size=2
+    - loader.eval_batch_size=8
+    - sampling.num_sample_batches=1
+    - sampling.steps=1
+    - eval.checkpoint_path=/home/ruizzhao/diffusion_baseline/checkpoints/official/duo/owt
+    - trainer.devices=1
+    - checkpointing.save_dir=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/duo/steps_1/upstream_hydra
+    - +wandb.offline=true
+    - eval.generated_samples_path=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/duo/steps_1/upstream_samples.json
+    - algo=duo_base
+    - algo.backbone=hf_dit
+    - sampling.predictor=ancestral
+    - sampling.noise_removal=ancestral
+    - eval.disable_ema=False
+  job:
+    name: dlb_pinned_upstream_main
+    chdir: true
+    override_dirname: +wandb.offline=true,algo.backbone=hf_dit,algo=duo_base,checkpointing.save_dir=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/duo/steps_1/upstream_hydra,data=openwebtext-split,eval.checkpoint_path=/home/ruizzhao/diffusion_baseline/checkpoints/official/duo/owt,eval.disable_ema=False,eval.generated_samples_path=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/duo/steps_1/upstream_samples.json,loader.batch_size=2,loader.eval_batch_size=8,mode=sample_eval,model.length=1024,model=small,sampling.noise_removal=ancestral,sampling.num_sample_batches=1,sampling.predictor=ancestral,sampling.steps=1,seed=42,trainer.devices=1
+    id: ???
+    num: ???
+    config_name: config
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /home/ruizzhao/diffusion_baseline
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /home/ruizzhao/diffusion_baseline/upstreams/duo/configs
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/duo/steps_1/upstream_hydra
+    choices:
+      algo: duo_base
+      prior: none
+      lr_scheduler: constant_warmup
+      noise: log-linear
+      strategy: ddp
+      model: small
+      data: openwebtext-split
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

results/smoke/samples/owt/duo/steps_16/stderr.log

@@ -0,0 +1,10 @@
+Seed set to 42
+/opt/anaconda3/envs/dlb-duo/lib/python3.12/site-packages/huggingface_hub/file_download.py:949: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+
  0%|          | 0/1 [00:00<?, ?it/s]/home/ruizzhao/.cache/huggingface/modules/transformers_modules/owt/model.py:544: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(
+/home/ruizzhao/.cache/huggingface/modules/transformers_modules/owt/model.py:161: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(enabled=False):
+/home/ruizzhao/.cache/huggingface/modules/transformers_modules/owt/model.py:116: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(enabled=False):
+

results/smoke/samples/owt/duo/steps_16/upstream_hydra/dlb_pinned_upstream_main.log

@@ -0,0 +1,5 @@
+[2026-08-03 13:01:15,779][dlb_pinned_upstream_main][INFO] - Starting Sample Eval.
+[2026-08-03 14:44:17,673][dlb_pinned_upstream_main][INFO] - Starting Sample Eval.
+[2026-08-03 14:44:30,781][dlb_pinned_upstream_main][INFO] - Generative perplexity: 121.81356048583984
+[2026-08-03 14:44:30,781][dlb_pinned_upstream_main][INFO] - Sample entropy: 5.549579620361328
+[2026-08-03 14:44:30,782][dlb_pinned_upstream_main][INFO] - Samples saved at: /home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/duo/steps_16/upstream_samples.json

results/smoke/samples/owt/duo/steps_32/stdout.log

@@ -0,0 +1,173 @@
+CONFIG
+├── mode
+│   └── sample_eval                                                             
+├── seed
+│   └── 42                                                                      
+├── loader
+│   └── global_batch_size: 512                                                  
+│       eval_global_batch_size: 512                                             
+│       batch_size: 2                                                           
+│       eval_batch_size: 8                                                      
+│       num_workers: 512                                                        
+│       pin_memory: true                                                        
+│                                                                               
+├── sampling
+│   └── predictor: ancestral                                                    
+│       steps: 32                                                               
+│       noise_removal: ancestral                                                
+│       use_float64: true                                                       
+│       p_nucleus: 1.0                                                          
+│       num_sample_batches: 1                                                   
+│       num_sample_log: 2                                                       
+│       semi_ar: false                                                          
+│       stride_length: 1                                                        
+│       num_strides: 1                                                          
+│       guid_weight: null                                                       
+│       psi:                                                                    
+│         time_profile: linear                                                  
+│         high_mode: pure-posterior                                             
+│         middle_mode: pure-posterior                                           
+│         low_mode: pure-posterior                                              
+│         high_frac: 0.2                                                        
+│         middle_frac: 0.6                                                      
+│                                                                               
+├── training
+│   └── ema: 0.9999                                                             
+│       antithetic_sampling: true                                               
+│       importance_sampling: false                                              
+│       sampling_eps: 0.001                                                     
+│       change_of_variables: false                                              
+│       loss_precision: bf16                                                    
+│       finetune_path: ''                                                       
+│       class_dropout_p: 0.1                                                    
+│                                                                               
+├── eval
+│   └── checkpoint_path: /home/ruizzhao/diffusion_baseline/checkpoints/official/
+│       disable_ema: false                                                      
+│       compute_generative_perplexity: false                                    
+│       perplexity_batch_size: 8                                                
+│       compute_perplexity_on_sanity: false                                     
+│       gen_ppl_eval_model_name_or_path: gpt2-large                             
+│       generate_samples: true                                                  
+│       generated_samples_path: /home/ruizzhao/diffusion_baseline/results/smoke/
+│                                                                               
+├── optim
+│   └── weight_decay: 0                                                         
+│       lr: 0.0003                                                              
+│       beta1: 0.9                                                              
+│       beta2: 0.999                                                            
+│       eps: 1.0e-08                                                            
+│                                                                               
+├── trainer
+│   └── _target_: lightning.Trainer                                             
+│       accelerator: cuda                                                       
+│       num_nodes: 1                                                            
+│       devices: 1                                                              
+│       accumulate_grad_batches: 256                                            
+│       gradient_clip_val: 1.0                                                  
+│       precision: bf16                                                         
+│       num_sanity_val_steps: 2                                                 
+│       max_steps: 1000000                                                      
+│       log_every_n_steps: 100                                                  
+│       limit_train_batches: 1.0                                                
+│       limit_val_batches: 1.0                                                  
+│       val_check_interval: 5000                                                
+│       check_val_every_n_epoch: null                                           
+│                                                                               
+├── wandb
+│   └── project: duo                                                            
+│       group: null                                                             
+│       job_type: null                                                          
+│       name: null                                                              
+│       id: None_42                                                             
+│       tags:                                                                   
+│       - log-linear                                                            
+│       - openwebtext-train                                                     
+│       - openwebtext-valid                                                     
+│       - duo_base                                                              
+│       offline: true                                                           
+│                                                                               
+├── checkpointing
+│   └── save_dir: /home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/du
+│       resume_from_ckpt: true                                                  
+│       resume_ckpt_path: /home/ruizzhao/diffusion_baseline/results/smoke/sample
+│                                                                               
+├── callbacks
+│   └── checkpoint_every_n_steps:                                               
+│         _target_: lightning.pytorch.callbacks.ModelCheckpoint                 
+│         save_top_k: -1                                                        
+│         save_last: true                                                       
+│         dirpath: /home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/d
+│         verbose: true                                                         
+│         auto_insert_metric_name: false                                        
+│         every_n_train_steps: 500                                              
+│       checkpoint_monitor:                                                     
+│         _target_: lightning.pytorch.callbacks.ModelCheckpoint                 
+│         monitor: val/nll                                                      
+│         mode: min                                                             
+│         save_top_k: 1                                                         
+│         save_last: false                                                      
+│         dirpath: /home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/d
+│         filename: best                                                        
+│         auto_insert_metric_name: false                                        
+│         verbose: true                                                         
+│       learning_rate_monitor:                                                  
+│         _target_: lightning.pytorch.callbacks.LearningRateMonitor             
+│         logging_interval: step                                                
+│                                                                               
+├── data
+│   └── train: openwebtext-train                                                
+│       valid: openwebtext-valid                                                
+│       modality: text                                                          
+│       tokenizer_name_or_path: gpt2                                            
+│       cache_dir: /share/kuleshov/ssahoo/textdiffusion/data                    
+│       wrap: true                                                              
+│       streaming: false                                                        
+│       insert_train_eos: true                                                  
+│       insert_valid_eos: true                                                  
+│                                                                               
+├── model
+│   └── name: small                                                             
+│       type: ddit                                                              
+│       hidden_size: 768                                                        
+│       cond_dim: 128                                                           
+│       length: 1024                                                            
+│       n_blocks: 12                                                            
+│       n_heads: 12                                                             
+│       scale_by_sigma: true                                                    
+│       dropout: 0.1                                                            
+│       tie_word_embeddings: false                                              
+│       vocab_lookup: true                                                      
+│                                                                               
+├── strategy
+│   └── _target_: lightning.pytorch.strategies.DDPStrategy                      
+│       find_unused_parameters: false                                           
+│                                                                               
+├── noise
+│   └── type: log-linear                                                        
+│       eps: 0.001                                                              
+│                                                                               
+├── lr_scheduler
+│   └── _target_: transformers.get_constant_schedule_with_warmup                
+│       num_warmup_steps: 2500                                                  
+│                                                                               
+├── prior
+│   └── type: none                                                              
+│       latent_width: 0                                                         
+│       latent_height: 0                                                        
+│                                                                               
+└── algo
+    └── name: duo_base                                                          
+        backbone: hf_dit                                                        
+        parameterization: mean                                                  
+        time_conditioning: true                                                 
+        T: 0                                                                    
+        subs_masking: false                                                     
+        causal_attention: false                                                 
+        ignore_bos: false                                                       
+        loss_type: elbo                                                         
+                                                                                
+[2026-08-03 14:44:19,355][dlb_pinned_upstream_main][INFO] - Starting Sample Eval.
+[2026-08-03 14:44:31,543][dlb_pinned_upstream_main][INFO] - Generative perplexity: 107.36914825439453
+[2026-08-03 14:44:31,543][dlb_pinned_upstream_main][INFO] - Sample entropy: 5.563689231872559
+[2026-08-03 14:44:31,544][dlb_pinned_upstream_main][INFO] - Samples saved at: /home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/duo/steps_32/upstream_samples.json

results/smoke/samples/owt/duo/steps_4/run_metadata.json

@@ -0,0 +1,78 @@
+{
+  "checkpoint_lock_id": "duo_owt_hf:3549a2eff7c0b0ccceca8d69a22baf7795fd9843fa9fcad41a8ddc186b9892f0:all",
+  "checkpoint_selection": {
+    "path": null,
+    "resource": "duo_owt_hf",
+    "teacher_family": null
+  },
+  "checkpoint_sha256": "d6f9b8ea4d06bad6be60d8c3a49284af44f3b38e9bc03b351c6c371239809785",
+  "checkpoint_teacher_family": "uniform_duo",
+  "command": [
+    "/opt/anaconda3/envs/dlb-duo/bin/python",
+    "-u",
+    "-m",
+    "dlb.adapters.capture",
+    "--upstream-entrypoint=/home/ruizzhao/diffusion_baseline/upstreams/duo/main.py",
+    "--capture-path=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/duo/steps_4/upstream_token_ids.json",
+    "--",
+    "mode=sample_eval",
+    "seed=42",
+    "data=openwebtext-split",
+    "model=small",
+    "model.length=1024",
+    "loader.batch_size=2",
+    "loader.eval_batch_size=8",
+    "sampling.num_sample_batches=1",
+    "sampling.steps=4",
+    "eval.checkpoint_path=/home/ruizzhao/diffusion_baseline/checkpoints/official/duo/owt",
+    "trainer.devices=1",
+    "hydra.run.dir=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/duo/steps_4/upstream_hydra",
+    "checkpointing.save_dir=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/duo/steps_4/upstream_hydra",
+    "+wandb.offline=true",
+    "eval.generated_samples_path=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/duo/steps_4/upstream_samples.json",
+    "algo=duo_base",
+    "algo.backbone=hf_dit",
+    "sampling.predictor=ancestral",
+    "sampling.noise_removal=ancestral",
+    "eval.disable_ema=False"
+  ],
+  "command_sha256": "11bb394e43fb812713407e65c81d7aa7829faf782c5ac2cf73488ff11c6f0062",
+  "config_sha256": "206d75b3b7cf3fb35844c053ae552537d7471b93bf29756dafc1ac616b096994",
+  "environment": {
+    "cuda_version": null,
+    "cuda_visible_devices": "3",
+    "hostname": "DS210029",
+    "platform": "Linux-5.15.0-186-generic-x86_64-with-glibc2.35",
+    "python": "3.12.13",
+    "torch_version": "2.5.1"
+  },
+  "finished_at": "2026-08-03T06:44:14.789407Z",
+  "identity": {
+    "adapter_identity": "dlb.adapters.duo:v1|dlb.adapters.duo:DuoAdapter|8878182eb10d7d1e3f18219cb62c2153084842fdb0ab532b1a3c7df0e10da3c6",
+    "checkpoint_lock_id": "duo_owt_hf:3549a2eff7c0b0ccceca8d69a22baf7795fd9843fa9fcad41a8ddc186b9892f0:all",
+    "checkpoint_selection": {
+      "path": null,
+      "resource": "duo_owt_hf",
+      "teacher_family": null
+    },
+    "checkpoint_sha256": "d6f9b8ea4d06bad6be60d8c3a49284af44f3b38e9bc03b351c6c371239809785",
+    "checkpoint_teacher_family": "uniform_duo",
+    "command_sha256": "11bb394e43fb812713407e65c81d7aa7829faf782c5ac2cf73488ff11c6f0062",
+    "config_sha256": "206d75b3b7cf3fb35844c053ae552537d7471b93bf29756dafc1ac616b096994",
+    "dataset_id": "owt",
+    "device": null,
+    "environment": "dlb-duo",
+    "model_id": "duo",
+    "results_root": "/home/ruizzhao/diffusion_baseline/results/smoke",
+    "run_id": "duo-owt-steps-4",
+    "sample_count": 1,
+    "seed": 42,
+    "source_sha256": "7c9b498f5b717de064d6fad7e2509c866e6cb620",
+    "step_count": 4
+  },
+  "source_sha256": "7c9b498f5b717de064d6fad7e2509c866e6cb620",
+  "started_at": "2026-08-03T06:44:02.198538Z",
+  "status": "succeeded",
+  "stderr_log": "/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/duo/steps_4/stderr.log",
+  "stdout_log": "/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/duo/steps_4/stdout.log"
+}

results/smoke/samples/owt/duo/steps_4/stdout.log

@@ -0,0 +1,173 @@
+CONFIG
+├── mode
+│   └── sample_eval                                                             
+├── seed
+│   └── 42                                                                      
+├── loader
+│   └── global_batch_size: 512                                                  
+│       eval_global_batch_size: 512                                             
+│       batch_size: 2                                                           
+│       eval_batch_size: 8                                                      
+│       num_workers: 512                                                        
+│       pin_memory: true                                                        
+│                                                                               
+├── sampling
+│   └── predictor: ancestral                                                    
+│       steps: 4                                                                
+│       noise_removal: ancestral                                                
+│       use_float64: true                                                       
+│       p_nucleus: 1.0                                                          
+│       num_sample_batches: 1                                                   
+│       num_sample_log: 2                                                       
+│       semi_ar: false                                                          
+│       stride_length: 1                                                        
+│       num_strides: 1                                                          
+│       guid_weight: null                                                       
+│       psi:                                                                    
+│         time_profile: linear                                                  
+│         high_mode: pure-posterior                                             
+│         middle_mode: pure-posterior                                           
+│         low_mode: pure-posterior                                              
+│         high_frac: 0.2                                                        
+│         middle_frac: 0.6                                                      
+│                                                                               
+├── training
+│   └── ema: 0.9999                                                             
+│       antithetic_sampling: true                                               
+│       importance_sampling: false                                              
+│       sampling_eps: 0.001                                                     
+│       change_of_variables: false                                              
+│       loss_precision: bf16                                                    
+│       finetune_path: ''                                                       
+│       class_dropout_p: 0.1                                                    
+│                                                                               
+├── eval
+│   └── checkpoint_path: /home/ruizzhao/diffusion_baseline/checkpoints/official/
+│       disable_ema: false                                                      
+│       compute_generative_perplexity: false                                    
+│       perplexity_batch_size: 8                                                
+│       compute_perplexity_on_sanity: false                                     
+│       gen_ppl_eval_model_name_or_path: gpt2-large                             
+│       generate_samples: true                                                  
+│       generated_samples_path: /home/ruizzhao/diffusion_baseline/results/smoke/
+│                                                                               
+├── optim
+│   └── weight_decay: 0                                                         
+│       lr: 0.0003                                                              
+│       beta1: 0.9                                                              
+│       beta2: 0.999                                                            
+│       eps: 1.0e-08                                                            
+│                                                                               
+├── trainer
+│   └── _target_: lightning.Trainer                                             
+│       accelerator: cuda                                                       
+│       num_nodes: 1                                                            
+│       devices: 1                                                              
+│       accumulate_grad_batches: 256                                            
+│       gradient_clip_val: 1.0                                                  
+│       precision: bf16                                                         
+│       num_sanity_val_steps: 2                                                 
+│       max_steps: 1000000                                                      
+│       log_every_n_steps: 100                                                  
+│       limit_train_batches: 1.0                                                
+│       limit_val_batches: 1.0                                                  
+│       val_check_interval: 5000                                                
+│       check_val_every_n_epoch: null                                           
+│                                                                               
+├── wandb
+│   └── project: duo                                                            
+│       group: null                                                             
+│       job_type: null                                                          
+│       name: null                                                              
+│       id: None_42                                                             
+│       tags:                                                                   
+│       - log-linear                                                            
+│       - openwebtext-train                                                     
+│       - openwebtext-valid                                                     
+│       - duo_base                                                              
+│       offline: true                                                           
+│                                                                               
+├── checkpointing
+│   └── save_dir: /home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/du
+│       resume_from_ckpt: true                                                  
+│       resume_ckpt_path: /home/ruizzhao/diffusion_baseline/results/smoke/sample
+│                                                                               
+├── callbacks
+│   └── checkpoint_every_n_steps:                                               
+│         _target_: lightning.pytorch.callbacks.ModelCheckpoint                 
+│         save_top_k: -1                                                        
+│         save_last: true                                                       
+│         dirpath: /home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/d
+│         verbose: true                                                         
+│         auto_insert_metric_name: false                                        
+│         every_n_train_steps: 500                                              
+│       checkpoint_monitor:                                                     
+│         _target_: lightning.pytorch.callbacks.ModelCheckpoint                 
+│         monitor: val/nll                                                      
+│         mode: min                                                             
+│         save_top_k: 1                                                         
+│         save_last: false                                                      
+│         dirpath: /home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/d
+│         filename: best                                                        
+│         auto_insert_metric_name: false                                        
+│         verbose: true                                                         
+│       learning_rate_monitor:                                                  
+│         _target_: lightning.pytorch.callbacks.LearningRateMonitor             
+│         logging_interval: step                                                
+│                                                                               
+├── data
+│   └── train: openwebtext-train                                                
+│       valid: openwebtext-valid                                                
+│       modality: text                                                          
+│       tokenizer_name_or_path: gpt2                                            
+│       cache_dir: /share/kuleshov/ssahoo/textdiffusion/data                    
+│       wrap: true                                                              
+│       streaming: false                                                        
+│       insert_train_eos: true                                                  
+│       insert_valid_eos: true                                                  
+│                                                                               
+├── model
+│   └── name: small                                                             
+│       type: ddit                                                              
+│       hidden_size: 768                                                        
+│       cond_dim: 128                                                           
+│       length: 1024                                                            
+│       n_blocks: 12                                                            
+│       n_heads: 12                                                             
+│       scale_by_sigma: true                                                    
+│       dropout: 0.1                                                            
+│       tie_word_embeddings: false                                              
+│       vocab_lookup: true                                                      
+│                                                                               
+├── strategy
+│   └── _target_: lightning.pytorch.strategies.DDPStrategy                      
+│       find_unused_parameters: false                                           
+│                                                                               
+├── noise
+│   └── type: log-linear                                                        
+│       eps: 0.001                                                              
+│                                                                               
+├── lr_scheduler
+│   └── _target_: transformers.get_constant_schedule_with_warmup                
+│       num_warmup_steps: 2500                                                  
+│                                                                               
+├── prior
+│   └── type: none                                                              
+│       latent_width: 0                                                         
+│       latent_height: 0                                                        
+│                                                                               
+└── algo
+    └── name: duo_base                                                          
+        backbone: hf_dit                                                        
+        parameterization: mean                                                  
+        time_conditioning: true                                                 
+        T: 0                                                                    
+        subs_masking: false                                                     
+        causal_attention: false                                                 
+        ignore_bos: false                                                       
+        loss_type: elbo                                                         
+                                                                                
+[2026-08-03 14:44:05,887][dlb_pinned_upstream_main][INFO] - Starting Sample Eval.
+[2026-08-03 14:44:13,672][dlb_pinned_upstream_main][INFO] - Generative perplexity: 496.36834716796875
+[2026-08-03 14:44:13,672][dlb_pinned_upstream_main][INFO] - Sample entropy: 5.506994724273682
+[2026-08-03 14:44:13,673][dlb_pinned_upstream_main][INFO] - Samples saved at: /home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/duo/steps_4/upstream_samples.json

results/smoke/samples/owt/flm/steps_1024/upstream_hydra/.hydra/hydra.yaml

@@ -0,0 +1,182 @@
+hydra:
+  run:
+    dir: /home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/flm/steps_1024/upstream_hydra
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.run.dir=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/flm/steps_1024/upstream_hydra
+    - hydra.mode=RUN
+    task:
+    - mode=sample_eval
+    - seed=42
+    - data=openwebtext-split
+    - model=small
+    - model.length=1024
+    - loader.batch_size=2
+    - loader.eval_batch_size=16
+    - sampling.num_sample_batches=1
+    - sampling.steps=1024
+    - eval.checkpoint_path=/home/ruizzhao/diffusion_baseline/checkpoints/official/flm_ckpt/owt/owt_flm.ckpt
+    - trainer.devices=1
+    - checkpointing.save_dir=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/flm/steps_1024/upstream_hydra
+    - +wandb.offline=true
+    - eval.generated_samples_path=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/flm/steps_1024/upstream_samples.json
+    - algo=flm
+    - algo.backbone=dit
+    - eval.disable_ema=False
+    - sampling.solver=euler
+    - algo.double_temb=False
+    - sampling.gamma=0.0
+  job:
+    name: dlb_pinned_upstream_main
+    chdir: true
+    override_dirname: +wandb.offline=true,algo.backbone=dit,algo.double_temb=False,algo=flm,checkpointing.save_dir=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/flm/steps_1024/upstream_hydra,data=openwebtext-split,eval.checkpoint_path=/home/ruizzhao/diffusion_baseline/checkpoints/official/flm_ckpt/owt/owt_flm.ckpt,eval.disable_ema=False,eval.generated_samples_path=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/flm/steps_1024/upstream_samples.json,loader.batch_size=2,loader.eval_batch_size=16,mode=sample_eval,model.length=1024,model=small,sampling.gamma=0.0,sampling.num_sample_batches=1,sampling.solver=euler,sampling.steps=1024,seed=42,trainer.devices=1
+    id: ???
+    num: ???
+    config_name: config
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /home/ruizzhao/diffusion_baseline
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /home/ruizzhao/diffusion_baseline/upstreams/flm/configs
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/flm/steps_1024/upstream_hydra
+    choices:
+      algo: flm
+      prior: none
+      lr_scheduler: constant_warmup
+      noise: log-linear
+      strategy: ddp
+      model: small
+      data: openwebtext-split
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

results/smoke/samples/owt/flm/steps_32/upstream_hydra/.hydra/hydra.yaml

@@ -0,0 +1,182 @@
+hydra:
+  run:
+    dir: /home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/flm/steps_32/upstream_hydra
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.run.dir=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/flm/steps_32/upstream_hydra
+    - hydra.mode=RUN
+    task:
+    - mode=sample_eval
+    - seed=42
+    - data=openwebtext-split
+    - model=small
+    - model.length=1024
+    - loader.batch_size=2
+    - loader.eval_batch_size=16
+    - sampling.num_sample_batches=1
+    - sampling.steps=32
+    - eval.checkpoint_path=/home/ruizzhao/diffusion_baseline/checkpoints/official/flm_ckpt/owt/owt_flm.ckpt
+    - trainer.devices=1
+    - checkpointing.save_dir=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/flm/steps_32/upstream_hydra
+    - +wandb.offline=true
+    - eval.generated_samples_path=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/flm/steps_32/upstream_samples.json
+    - algo=flm
+    - algo.backbone=dit
+    - eval.disable_ema=False
+    - sampling.solver=euler
+    - algo.double_temb=False
+    - sampling.gamma=0.0
+  job:
+    name: dlb_pinned_upstream_main
+    chdir: true
+    override_dirname: +wandb.offline=true,algo.backbone=dit,algo.double_temb=False,algo=flm,checkpointing.save_dir=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/flm/steps_32/upstream_hydra,data=openwebtext-split,eval.checkpoint_path=/home/ruizzhao/diffusion_baseline/checkpoints/official/flm_ckpt/owt/owt_flm.ckpt,eval.disable_ema=False,eval.generated_samples_path=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/flm/steps_32/upstream_samples.json,loader.batch_size=2,loader.eval_batch_size=16,mode=sample_eval,model.length=1024,model=small,sampling.gamma=0.0,sampling.num_sample_batches=1,sampling.solver=euler,sampling.steps=32,seed=42,trainer.devices=1
+    id: ???
+    num: ???
+    config_name: config
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.3'
+    cwd: /home/ruizzhao/diffusion_baseline
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /home/ruizzhao/diffusion_baseline/upstreams/flm/configs
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/flm/steps_32/upstream_hydra
+    choices:
+      algo: flm
+      prior: none
+      lr_scheduler: constant_warmup
+      noise: log-linear
+      strategy: ddp
+      model: small
+      data: openwebtext-split
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

results/smoke/samples/owt/fmlm/steps_16/request.json

@@ -0,0 +1,90 @@
+{
+  "command": [
+    "/opt/anaconda3/envs/dlb-flm/bin/python",
+    "-u",
+    "-m",
+    "dlb.adapters.capture",
+    "--upstream-entrypoint=/home/ruizzhao/diffusion_baseline/upstreams/flm/main.py",
+    "--capture-path=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/fmlm/steps_16/upstream_token_ids.json",
+    "--",
+    "mode=sample_eval",
+    "seed=42",
+    "data=openwebtext-split",
+    "model=small",
+    "model.length=1024",
+    "loader.batch_size=2",
+    "loader.eval_batch_size=16",
+    "sampling.num_sample_batches=1",
+    "sampling.steps=16",
+    "eval.checkpoint_path=/home/ruizzhao/diffusion_baseline/checkpoints/official/fmlm_ckpt/owt/owt_fmlm.ckpt",
+    "trainer.devices=1",
+    "hydra.run.dir=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/fmlm/steps_16/upstream_hydra",
+    "checkpointing.save_dir=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/fmlm/steps_16/upstream_hydra",
+    "+wandb.offline=true",
+    "eval.generated_samples_path=/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/fmlm/steps_16/upstream_samples.json",
+    "algo=fmlm",
+    "algo.backbone=dit",
+    "eval.disable_ema=False",
+    "sampling.solver=euler",
+    "algo.double_temb=True",
+    "algo.learnable_loss_weighting=False",
+    "sampling.gamma=1.0"
+  ],
+  "identity": {
+    "adapter_identity": "dlb.adapters.flm:v2|dlb.adapters.flm:FLMAdapter|c3a6a4573200fecf38a47fc6e7bc78254ca362fff3b96079dc2efa9f8e27aff0",
+    "checkpoint_lock_id": "recipe:fmlm_owt_official_ckpt:3549a2eff7c0b0ccceca8d69a22baf7795fd9843fa9fcad41a8ddc186b9892f0",
+    "checkpoint_selection": {
+      "output": "checkpoints/official/fmlm_ckpt/owt",
+      "recipe": "fmlm_owt_official_ckpt",
+      "sampling_checkpoint": "owt_fmlm.ckpt",
+      "sampling_config": null,
+      "source": "flm",
+      "source_commit": "a1918d5164e5038e37d0b7a4fb2010ce75b863b3",
+      "teacher_adapter": null,
+      "teacher_family": "continuous_flm"
+    },
+    "checkpoint_sha256": "a5ca77b77a26dec9746f627858d42c19803d11c1e932dcf14a3ce80f6bbbbd76",
+    "checkpoint_teacher_family": "continuous_flm",
+    "command_sha256": "d13885d4326c17c9ddd2fe9519ba33c45253f2eeb28a811ebaa44f7af1fa5632",
+    "config_sha256": "206d75b3b7cf3fb35844c053ae552537d7471b93bf29756dafc1ac616b096994",
+    "dataset_id": "owt",
+    "device": null,
+    "environment": "dlb-flm",
+    "model_id": "fmlm",
+    "results_root": "/home/ruizzhao/diffusion_baseline/results/smoke",
+    "run_id": "fmlm-owt-steps-16",
+    "sample_count": 1,
+    "seed": 42,
+    "source_sha256": "a1918d5164e5038e37d0b7a4fb2010ce75b863b3",
+    "step_count": 16
+  },
+  "request": {
+    "adapter_identity": "dlb.adapters.flm:v2|dlb.adapters.flm:FLMAdapter|c3a6a4573200fecf38a47fc6e7bc78254ca362fff3b96079dc2efa9f8e27aff0",
+    "checkpoint_lock_id": "recipe:fmlm_owt_official_ckpt:3549a2eff7c0b0ccceca8d69a22baf7795fd9843fa9fcad41a8ddc186b9892f0",
+    "checkpoint_selection": {
+      "output": "checkpoints/official/fmlm_ckpt/owt",
+      "recipe": "fmlm_owt_official_ckpt",
+      "sampling_checkpoint": "owt_fmlm.ckpt",
+      "sampling_config": null,
+      "source": "flm",
+      "source_commit": "a1918d5164e5038e37d0b7a4fb2010ce75b863b3",
+      "teacher_adapter": null,
+      "teacher_family": "continuous_flm"
+    },
+    "checkpoint_sha256": "a5ca77b77a26dec9746f627858d42c19803d11c1e932dcf14a3ce80f6bbbbd76",
+    "checkpoint_teacher_family": "continuous_flm",
+    "command": null,
+    "config_sha256": "206d75b3b7cf3fb35844c053ae552537d7471b93bf29756dafc1ac616b096994",
+    "dataset_id": "owt",
+    "device": null,
+    "environment": "dlb-flm",
+    "model_id": "fmlm",
+    "results_root": "/home/ruizzhao/diffusion_baseline/results/smoke",
+    "run_id": "fmlm-owt-steps-16",
+    "sample_count": 1,
+    "seed": 42,
+    "source_sha256": "a1918d5164e5038e37d0b7a4fb2010ce75b863b3",
+    "step_count": 16
+  },
+  "started_at": "2026-08-03T06:48:06.847669Z"
+}

results/smoke/samples/owt/fmlm/steps_8/conversion_metadata.json

@@ -0,0 +1,31 @@
+{
+  "checkpoint_lock_id": "recipe:fmlm_owt_official_ckpt:3549a2eff7c0b0ccceca8d69a22baf7795fd9843fa9fcad41a8ddc186b9892f0",
+  "checkpoint_selection": {
+    "output": "checkpoints/official/fmlm_ckpt/owt",
+    "recipe": "fmlm_owt_official_ckpt",
+    "sampling_checkpoint": "owt_fmlm.ckpt",
+    "sampling_config": null,
+    "source": "flm",
+    "source_commit": "a1918d5164e5038e37d0b7a4fb2010ce75b863b3",
+    "teacher_adapter": null,
+    "teacher_family": "continuous_flm"
+  },
+  "checkpoint_sha256": "a5ca77b77a26dec9746f627858d42c19803d11c1e932dcf14a3ce80f6bbbbd76",
+  "exclude_from_latency": true,
+  "format": "upstream-generated-seqs-v1",
+  "generated_samples": 16,
+  "generation_seconds_sentinel": 0.0,
+  "generation_seconds_source": "unavailable_excluded_sentinel",
+  "requested_samples": 1,
+  "teacher_family": "continuous_flm",
+  "token_ids_source": "upstream",
+  "token_ids_transformation": {
+    "max_length": 1024,
+    "operation": "validated_exact_length"
+  },
+  "tokenizer": "gpt2",
+  "tokenizer_revision": "607a30d783dfa663caf39e06633721c8d4cfcd7e",
+  "trim_policy": "stable_prefix",
+  "trimmed_samples": 15,
+  "upstream_output": "/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/fmlm/steps_8/upstream_samples.json"
+}

results/smoke/samples/owt/mdlm_di4c/steps_4/request.json

@@ -0,0 +1,99 @@
+{
+  "command": [
+    "/opt/anaconda3/envs/dlb-di4c/bin/python",
+    "-B",
+    "-u",
+    "/home/ruizzhao/diffusion_baseline/adapters/sample_di4c.py",
+    "--upstream-root",
+    "/home/ruizzhao/diffusion_baseline/upstreams/di4c/sdtt",
+    "--checkpoint",
+    "/home/ruizzhao/diffusion_baseline/checkpoints/official/mdlm_di4c/owt/sdtt7-di4c2.ckpt",
+    "--config",
+    "/home/ruizzhao/diffusion_baseline/configs/sampling/di4c_mdlm_owt.yaml",
+    "--checkpoint-sha256",
+    "4af61572dc82fe012ce395a2e4e4be0585ea09076597ba1f7241ca196b97ecb0",
+    "--config-sha256",
+    "cf627f89b8dede9b25ac8e10407c3b7de203a76c4ba84100723287e571dcec34",
+    "--data-config",
+    "/home/ruizzhao/diffusion_baseline/artifacts/data.yaml",
+    "--downloads-manifest",
+    "/home/ruizzhao/diffusion_baseline/data/manifests/downloads.json",
+    "--tokenizer-snapshot",
+    "/home/ruizzhao/diffusion_baseline/data/raw/huggingface/hub/models--gpt2/snapshots/607a30d783dfa663caf39e06633721c8d4cfcd7e",
+    "--output",
+    "/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/mdlm_di4c/steps_4/upstream_token_ids.json",
+    "--sample-count",
+    "1",
+    "--batch-size",
+    "1",
+    "--num-steps",
+    "4",
+    "--seq-len",
+    "1024",
+    "--seed",
+    "42",
+    "--sampler",
+    "ancestral",
+    "--teacher-family",
+    "masked_mdlm",
+    "--dataset",
+    "owt",
+    "--offline",
+    "true"
+  ],
+  "identity": {
+    "adapter_identity": "dlb.adapters.di4c:v3|dlb.adapters.di4c:Di4CAdapter|ba289424e1babd29c99b4bb95c008a934d000e1ddb9e76140f3963398bf02111",
+    "checkpoint_lock_id": "di4c_mdlm_owt_zenodo:3549a2eff7c0b0ccceca8d69a22baf7795fd9843fa9fcad41a8ddc186b9892f0:all",
+    "checkpoint_selection": {
+      "path": null,
+      "resource": "di4c_mdlm_owt_zenodo",
+      "sampling_config": "configs/sampling/di4c_mdlm_owt.yaml",
+      "sampling_config_sha256": "cf627f89b8dede9b25ac8e10407c3b7de203a76c4ba84100723287e571dcec34",
+      "sampling_config_source": "project",
+      "sampling_config_source_commit": "ac61ff9fe8e85120f9e1d2a8c5a332f8b8353dd3",
+      "teacher_family": null
+    },
+    "checkpoint_sha256": "e51269cd095bb1ed1b8771e4b5c8470d22b01fa7fa0652dce8eff18fcb064f8c",
+    "checkpoint_teacher_family": "masked_mdlm",
+    "command_sha256": "4448eae936eaa6891928f885fffd3c66db53aee4444dc9cf07ae6e2139d1863b",
+    "config_sha256": "206d75b3b7cf3fb35844c053ae552537d7471b93bf29756dafc1ac616b096994",
+    "dataset_id": "owt",
+    "device": null,
+    "environment": "dlb-di4c",
+    "model_id": "mdlm_di4c",
+    "results_root": "/home/ruizzhao/diffusion_baseline/results/smoke",
+    "run_id": "mdlm_di4c-owt-steps-4",
+    "sample_count": 1,
+    "seed": 42,
+    "source_sha256": "ac61ff9fe8e85120f9e1d2a8c5a332f8b8353dd3",
+    "step_count": 4
+  },
+  "request": {
+    "adapter_identity": "dlb.adapters.di4c:v3|dlb.adapters.di4c:Di4CAdapter|ba289424e1babd29c99b4bb95c008a934d000e1ddb9e76140f3963398bf02111",
+    "checkpoint_lock_id": "di4c_mdlm_owt_zenodo:3549a2eff7c0b0ccceca8d69a22baf7795fd9843fa9fcad41a8ddc186b9892f0:all",
+    "checkpoint_selection": {
+      "path": null,
+      "resource": "di4c_mdlm_owt_zenodo",
+      "sampling_config": "configs/sampling/di4c_mdlm_owt.yaml",
+      "sampling_config_sha256": "cf627f89b8dede9b25ac8e10407c3b7de203a76c4ba84100723287e571dcec34",
+      "sampling_config_source": "project",
+      "sampling_config_source_commit": "ac61ff9fe8e85120f9e1d2a8c5a332f8b8353dd3",
+      "teacher_family": null
+    },
+    "checkpoint_sha256": "e51269cd095bb1ed1b8771e4b5c8470d22b01fa7fa0652dce8eff18fcb064f8c",
+    "checkpoint_teacher_family": "masked_mdlm",
+    "command": null,
+    "config_sha256": "206d75b3b7cf3fb35844c053ae552537d7471b93bf29756dafc1ac616b096994",
+    "dataset_id": "owt",
+    "device": null,
+    "environment": "dlb-di4c",
+    "model_id": "mdlm_di4c",
+    "results_root": "/home/ruizzhao/diffusion_baseline/results/smoke",
+    "run_id": "mdlm_di4c-owt-steps-4",
+    "sample_count": 1,
+    "seed": 42,
+    "source_sha256": "ac61ff9fe8e85120f9e1d2a8c5a332f8b8353dd3",
+    "step_count": 4
+  },
+  "started_at": "2026-08-03T06:48:33.478108Z"
+}

results/smoke/samples/owt/mdlm_di4c/steps_4/stderr.log

@@ -0,0 +1,9 @@
+/opt/anaconda3/envs/dlb-di4c/lib/python3.10/site-packages/lightning/fabric/__init__.py:41: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.
+Traceback (most recent call last):
+  File "/home/ruizzhao/diffusion_baseline/adapters/sample_di4c.py", line 124, in <module>
+    raise SystemExit(main())
+  File "/home/ruizzhao/diffusion_baseline/adapters/sample_di4c.py", line 85, in main
+    validate_embedded_config(config, embedded_config)
+  File "/home/ruizzhao/diffusion_baseline/adapters/_distilled_runtime.py", line 246, in validate_embedded_config
+    raise ValueError("Lightning checkpoint does not embed its training config")
+ValueError: Lightning checkpoint does not embed its training config

results/smoke/samples/owt/mdlm_sdtt/steps_1/conversion_metadata.json

@@ -0,0 +1,30 @@
+{
+  "checkpoint_lock_id": "sdtt_owt_hf:3549a2eff7c0b0ccceca8d69a22baf7795fd9843fa9fcad41a8ddc186b9892f0:all",
+  "checkpoint_selection": {
+    "path": null,
+    "resource": "sdtt_owt_hf",
+    "sampling_config": "config.json",
+    "sampling_config_sha256": null,
+    "sampling_config_source": "resource",
+    "sampling_config_source_commit": null,
+    "teacher_family": null
+  },
+  "checkpoint_sha256": "79eda51b8c8e8877b750b24a2276e240eeb2fd4230f369c99d71fc33bb113628",
+  "exclude_from_latency": true,
+  "format": "dlb-upstream-token-capture-v1",
+  "generated_samples": 1,
+  "generation_seconds_sentinel": 0.0,
+  "generation_seconds_source": "unavailable_excluded_sentinel",
+  "requested_samples": 1,
+  "teacher_family": "masked_mdlm",
+  "token_ids_source": "upstream",
+  "token_ids_transformation": {
+    "max_length": 1024,
+    "operation": "validated_exact_length"
+  },
+  "tokenizer": "gpt2",
+  "tokenizer_revision": "607a30d783dfa663caf39e06633721c8d4cfcd7e",
+  "trim_policy": "none_exact_count",
+  "trimmed_samples": 0,
+  "upstream_output": "/home/ruizzhao/diffusion_baseline/results/smoke/samples/owt/mdlm_sdtt/steps_1/upstream_token_ids.json"
+}

results/smoke/samples/owt/mdlm_sdtt/steps_1/stderr.log

@@ -0,0 +1 @@
+/opt/anaconda3/envs/dlb-sdtt/lib/python3.10/site-packages/lightning/fabric/__init__.py:41: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.

results/smoke/samples/owt/mdlm_sdtt/steps_32/samples.jsonl

@@ -0,0 +1 @@
+{"generation_seconds":0.0,"sample_id":0,"seed":42,"text":"<|endoftext|> to be no guarantee that we will get it when we’ve done our research. My hope is that it will work, and we can eventually coax all people in this scheme into quitting,” he says. “By then, a young family was living in Mexico, in California, and after they found a cannabis camp in Tierra, I wanted to start helping. I had worked at an agency that runs the camps on bikes, so I wanted to be based in this area to help these young people, in turn, get into work for it, and ultimately make it available to people who are not in their family in the 18th grade or too late.”\n\nJonathan Hector, 23, who lives in San Francisco, says:\n\nFor years, he’s struggled to cope now. He’s been on a study tour, and soon discovered that in many different ways, it is difficult to find a time to reflect on my personal life choices, and to try to moderate the amount thoughts I have to take in relation to my medications. Also, a lot of people find it extremely difficult to watch their choices. There is a very large group of people who are most frustrated, they want to reduce the anxiety. In particular, they are not afraid to wait for a long time before they take any medication, and they are determined to avoid taking [their] medicine. They are determined to save money and try also to minimise the anxiety. Also, because these people are living here, they’re not taking recreational activities; they are taking family-oriented activities.\n\nSometime after he was commissioned, Hector received standard marijuana treatment and he decided to stop this treatment, for the last two years, because the doctors had ordered him he was not going to because he didn’t take his medication, or this would question the duration of the treatment. But he was satisfied, and so decided to take cannabis, only to do so again in two different groups of people.\n\nThe internet class was organised, and by then, he was travelling to Mexico, where a friend of mine offered to pay him for cannabis worth of medicinal cannabis, and for the extra money he wouldn’t take traditional, herbal, or even both, cannabis. The medicinal cannabis course was planned to start once, and this involved a range of different groups of respondents, and a small number of existing interviewees, but the group was not very distinct. Instead, there were a groups made of 17 randomly named people selected by one of their ethnic groups, where the participants stayed, and adapted their personal experiences. There were some Chinese participants who preferred traditional medicinal cannabis, some Japanese preferred cannabis, and by then there was a similar pool.\n\nFrom there, 37 key respondents were asked about their participation in the course. Participants were randomly allocated to proceed as if they were male to medical cannabis, according to those by nature were the most interested in cannabis, and in the early stages of treatment, tended to take their own medications at a lower rate.However, there is on ongoing research showing that older people – most often those who are unable to take cannabinoids – spend often a lot more time in the CBD. There are also a number of people who smoked while taking cannabis, and still have high rates of the disease, and considering more of those even smoked for their own benefit than the rest. However, there is only the possibility of transitioning an older patient to cannabinoid therapy, there is also the possibility of developing new therapies for cannabinoids that patients use as well.\n\n to the United States where Turkish drug minister Akbar Barış is expected to wait for guests for the use of his CBD at Istanbul Airport, Turkey, June 16, 2015. file photo.\n\nIf these findings are published, they can be applied to a variety of other diseases, some of the most important childhood neurological diseases. Thus, they will be able to develop substantially more effective and effective cannabinoid cures and therapy for both children and adolescents, and in their lives. Cannabis is also strongly associated with people who use them. Pakistan currently shows a very important proportion of those, and Iran, the largest segment of these, and this trend will most accelerate in the UK, as it will reduce language barriers and making it easier for cannabis treatment specialists. Researchers will to find a new way in reducing all access to drugs, and I’m hoping with this study, they can do that in other time.<|endoftext|>Joshua Shippuzzi’s Peeled Plant and Litter Girl\n\nThe saffron comes in varieties of edible plants and a variety of Each of the many cre the small plant can grow. The tart strawberry is larger than any plant, about three to four meters. Passing through the ripening process and into the leaves, surrounding the roots, energy is passed to the green and lemon. The berries are produced by alternating light and dark sources, such as green colours, sunlight, and wind. In the leaves, air can form sulfuric acid, rings<|endoftext|>","token_ids":[50256,284,307,645,9149,326,356,481,651,340,618,356,447,247,303,1760,674,2267,13,2011,2911,318,326,340,481,670,11,290,356,460,4191,37700,477,661,287,428,7791,656,34161,11,447,251,339,1139,13,564,250,3886,788,11,257,1862,1641,373,2877,287,5828,11,287,3442,11,290,706,484,1043,257,9499,1413,287,15917,430,11,314,2227,284,923,5742,13,314,550,3111,379,281,4086,326,4539,262,11665,319,16715,11,523,314,2227,284,307,1912,287,428,1989,284,1037,777,1862,661,11,287,1210,11,651,656,670,329,340,11,290,6165,787,340,1695,284,661,508,389,407,287,511,1641,287,262,1248,400,9559,393,1165,2739,13,447,251,198,198,30365,42621,11,2242,11,508,3160,287,2986,6033,11,1139,25,198,198,1890,812,11,339,447,247,82,11615,284,19271,783,13,679,447,247,82,587,319,257,2050,4205,11,290,2582,5071,326,287,867,1180,2842,11,340,318,2408,284,1064,257,640,284,4079,319,616,2614,1204,7747,11,290,284,1949,284,10768,262,2033,6066,314,423,284,1011,287,8695,284,616,17638,13,4418,11,257,1256,286,661,1064,340,4457,2408,284,2342,511,7747,13,1318,318,257,845,1588,1448,286,661,508,389,749,14718,11,484,765,284,4646,262,9751,13,554,1948,11,484,389,407,7787,284,4043,329,257,890,640,878,484,1011,597,14103,11,290,484,389,5295,284,3368,2263,685,24571,60,9007,13,1119,389,5295,284,3613,1637,290,1949,635,284,10356,786,262,9751,13,4418,11,780,777,661,389,2877,994,11,484,447,247,260,407,2263,18136,4568,26,484,389,2263,1641,12,17107,4568,13,198,198,50,47056,706,339,373,20462,11,42621,2722,3210,5727,3513,290,339,3066,284,2245,428,3513,11,329,262,938,734,812,11,780,262,7519,550,6149,683,339,373,407,1016,284,780,339,1422,447,247,83,1011,465,14103,11,393,428,561,1808,262,9478,286,262,3513,13,887,339,373,11378,11,290,523,3066,284,1011,9499,11,691,284,466,523,757,287,734,1180,2628,286,661,13,198,198,464,5230,1398,373,20325,11,290,416,788,11,339,373,16574,284,5828,11,810,257,1545,286,6164,4438,284,1414,683,329,9499,2861,286,30381,9499,11,290,329,262,3131,1637,339,3636,447,247,83,1011,4569,11,35867,11,393,772,1111,11,9499,13,383,30381,9499,1781,373,6027,284,923,1752,11,290,428,2950,257,2837,286,1180,2628,286,14502,11,290,257,1402,1271,286,4683,2720,2841,11,475,262,1448,373,407,845,7310,13,5455,11,612,547,257,2628,925,286,1596,15456,3706,661,6163,416,530,286,511,9450,2628,11,810,262,6809,9658,11,290,16573,511,2614,6461,13,1318,547,617,3999,6809,508,9871,4569,30381,9499,11,617,4960,9871,9499,11,290,416,788,612,373,257,2092,5933,13,198,198,4863,612,11,5214,1994,14502,547,1965,546,511,10270,287,262,1781,13,26122,547,15456,19171,284,5120,355,611,484,547,4257,284,3315,9499,11,1864,284,883,416,3450,547,262,749,4609,287,9499,11,290,287,262,1903,9539,286,3513,11,19960,284,1011,511,898,17638,379,257,2793,2494,13,4864,11,612,318,319,7044,2267,4478,326,4697,661,784,749,1690,883,508,389,5906,284,1011,46830,784,4341,1690,257,1256,517,640,287,262,27147,13,1318,389,635,257,1271,286,661,508,21603,981,2263,9499,11,290,991,423,1029,3965,286,262,4369,11,290,6402,517,286,883,772,21603,329,511,898,4414,621,262,1334,13,2102,11,612,318,691,262,5885,286,37005,281,4697,5827,284,48175,9102,11,612,318,635,262,5885,286,5922,649,29596,329,46830,326,3871,779,355,880,13,198,198,284,262,1578,1829,810,9663,2563,5342,9084,5657,2409,30102,46481,318,2938,284,4043,329,10650,329,262,779,286,465,27147,379,25299,12690,11,7137,11,2795,1467,11,1853,13,2393,4590,13,198,198,1532,777,6373,389,3199,11,484,460,307,5625,284,257,4996,286,584,10040,11,617,286,262,749,1593,9963,30232,10040,13,6660,11,484,481,307,1498,284,1205,13788,517,4050,290,4050,48175,50019,290,9102,329,1111,1751,290,25447,11,290,287,511,3160,13,28838,318,635,7634,3917,351,661,508,779,606,13,7648,3058,2523,257,845,1593,9823,286,883,11,290,4068,11,262,4387,10618,286,777,11,290,428,5182,481,749,22636,287,262,3482,11,355,340,481,4646,3303,14725,290,1642,340,4577,329,9499,3513,22447,13,26685,481,284,1064,257,649,835,287,8868,477,1895,284,5010,11,290,314,447,247,76,7725,351,428,2050,11,484,460,466,326,287,584,640,13,50256,47740,911,3974,4715,72,447,247,82,2631,18449,16561,290,406,1967,7430,198,198,464,473,487,1313,2058,287,20328,286,35988,6134,290,257,4996,286,5501,286,262,867,1126,262,1402,4618,460,1663,13,383,35842,41236,318,4025,621,597,4618,11,546,1115,284,1440,10700,13,46389,832,262,18344,3101,1429,290,656,262,5667,11,7346,262,11135,11,2568,318,3804,284,262,4077,290,18873,13,383,36322,389,4635,416,39623,1657,290,3223,4237,11,884,355,4077,18915,11,19606,11,290,2344,13,554,262,5667,11,1633,460,1296,37402,291,7408,11,13917,50256]}

results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_1/upstream_hydra/.hydra/config.yaml

@@ -0,0 +1,151 @@
+mode: sample_eval
+seed: 42
+loader:
+  global_batch_size: 512
+  eval_global_batch_size: ${.global_batch_size}
+  batch_size: 2
+  eval_batch_size: 1
+  num_workers: ${eval:"len(__import__('os').sched_getaffinity(0))"}
+  pin_memory: true
+sampling:
+  predictor: ancestral
+  steps: 1
+  noise_removal: ancestral
+  use_float64: true
+  p_nucleus: 1.0
+  num_sample_batches: 1
+  num_sample_log: 2
+  semi_ar: false
+  stride_length: 1
+  num_strides: 1
+  guid_weight: null
+  psi:
+    time_profile: linear
+    high_mode: pure-posterior
+    middle_mode: pure-posterior
+    low_mode: pure-posterior
+    high_frac: 0.2
+    middle_frac: 0.6
+training:
+  ema: 0.9999
+  antithetic_sampling: true
+  importance_sampling: false
+  sampling_eps: 0.001
+  change_of_variables: false
+  loss_precision: bf16
+  finetune_path: ''
+  class_dropout_p: 0.1
+eval:
+  checkpoint_path: /home/ruizzhao/diffusion_baseline/checkpoints/reference_reproduction/duo_dcd/lm1b/model.ckpt
+  disable_ema: false
+  compute_generative_perplexity: false
+  perplexity_batch_size: 8
+  compute_perplexity_on_sanity: false
+  gen_ppl_eval_model_name_or_path: gpt2-large
+  generate_samples: true
+  generated_samples_path: /home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_1/upstream_samples.json
+optim:
+  weight_decay: 0
+  lr: 0.0003
+  beta1: 0.9
+  beta2: 0.999
+  eps: 1.0e-08
+trainer:
+  _target_: lightning.Trainer
+  accelerator: cuda
+  num_nodes: 1
+  devices: 1
+  accumulate_grad_batches: ${div_up:${loader.global_batch_size}, ${eval:${trainer.devices}
+    * ${loader.batch_size} * ${trainer.num_nodes}}}
+  gradient_clip_val: 1.0
+  precision: bf16
+  num_sanity_val_steps: 2
+  max_steps: 1000000
+  log_every_n_steps: 100
+  limit_train_batches: 1.0
+  limit_val_batches: 1.0
+  val_check_interval: 5000
+  check_val_every_n_epoch: null
+wandb:
+  project: duo
+  group: null
+  job_type: null
+  name: null
+  id: ${.name}_${seed}
+  tags:
+  - ${noise.type}
+  - ${data.train}
+  - ${data.valid}
+  - ${algo.name}
+  offline: true
+checkpointing:
+  save_dir: /home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_1/upstream_hydra
+  resume_from_ckpt: true
+  resume_ckpt_path: ${.save_dir}/checkpoints/last.ckpt
+callbacks:
+  checkpoint_every_n_steps:
+    _target_: lightning.pytorch.callbacks.ModelCheckpoint
+    save_top_k: -1
+    save_last: true
+    dirpath: ${checkpointing.save_dir}/checkpoints
+    verbose: true
+    auto_insert_metric_name: false
+    every_n_train_steps: 500
+  checkpoint_monitor:
+    _target_: lightning.pytorch.callbacks.ModelCheckpoint
+    monitor: val/nll
+    mode: min
+    save_top_k: 1
+    save_last: false
+    dirpath: ${checkpointing.save_dir}/checkpoints
+    filename: best
+    auto_insert_metric_name: false
+    verbose: true
+  learning_rate_monitor:
+    _target_: lightning.pytorch.callbacks.LearningRateMonitor
+    logging_interval: step
+data:
+  train: lm1b
+  valid: lm1b
+  modality: text
+  tokenizer_name_or_path: bert-base-uncased
+  cache_dir: /share/kuleshov/ssahoo/textdiffusion/data
+  wrap: true
+  streaming: false
+  insert_train_eos: true
+  insert_valid_eos: true
+model:
+  name: small
+  type: ddit
+  hidden_size: 768
+  cond_dim: 128
+  length: 128
+  n_blocks: 12
+  n_heads: 12
+  scale_by_sigma: true
+  dropout: 0.1
+  tie_word_embeddings: false
+  vocab_lookup: true
+strategy:
+  _target_: lightning.pytorch.strategies.DDPStrategy
+  find_unused_parameters: false
+noise:
+  type: log-linear
+  eps: 0.001
+lr_scheduler:
+  _target_: transformers.get_constant_schedule_with_warmup
+  num_warmup_steps: 2500
+prior:
+  type: none
+  latent_width: 0
+  latent_height: 0
+algo:
+  name: duo_base
+  backbone: dit
+  parameterization: mean
+  time_conditioning: true
+  T: 0
+  subs_masking: false
+  causal_attention: false
+  ignore_bos: false
+  loss_type: elbo

results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_1/upstream_hydra/config_tree.txt

@@ -0,0 +1,169 @@
+CONFIG
+├── mode
+│   └── sample_eval                                                                                                                           
+├── seed
+│   └── 42                                                                                                                                    
+├── loader
+│   └── global_batch_size: 512                                                                                                                
+│       eval_global_batch_size: 512                                                                                                           
+│       batch_size: 2                                                                                                                         
+│       eval_batch_size: 1                                                                                                                    
+│       num_workers: 512                                                                                                                      
+│       pin_memory: true                                                                                                                      
+│                                                                                                                                             
+├── sampling
+│   └── predictor: ancestral                                                                                                                  
+│       steps: 1                                                                                                                              
+│       noise_removal: ancestral                                                                                                              
+│       use_float64: true                                                                                                                     
+│       p_nucleus: 1.0                                                                                                                        
+│       num_sample_batches: 1                                                                                                                 
+│       num_sample_log: 2                                                                                                                     
+│       semi_ar: false                                                                                                                        
+│       stride_length: 1                                                                                                                      
+│       num_strides: 1                                                                                                                        
+│       guid_weight: null                                                                                                                     
+│       psi:                                                                                                                                  
+│         time_profile: linear                                                                                                                
+│         high_mode: pure-posterior                                                                                                           
+│         middle_mode: pure-posterior                                                                                                         
+│         low_mode: pure-posterior                                                                                                            
+│         high_frac: 0.2                                                                                                                      
+│         middle_frac: 0.6                                                                                                                    
+│                                                                                                                                             
+├── training
+│   └── ema: 0.9999                                                                                                                           
+│       antithetic_sampling: true                                                                                                             
+│       importance_sampling: false                                                                                                            
+│       sampling_eps: 0.001                                                                                                                   
+│       change_of_variables: false                                                                                                            
+│       loss_precision: bf16                                                                                                                  
+│       finetune_path: ''                                                                                                                     
+│       class_dropout_p: 0.1                                                                                                                  
+│                                                                                                                                             
+├── eval
+│   └── checkpoint_path: /home/ruizzhao/diffusion_baseline/checkpoints/reference_reproduction/duo_dcd/lm1b/model.ckpt                         
+│       disable_ema: false                                                                                                                    
+│       compute_generative_perplexity: false                                                                                                  
+│       perplexity_batch_size: 8                                                                                                              
+│       compute_perplexity_on_sanity: false                                                                                                   
+│       gen_ppl_eval_model_name_or_path: gpt2-large                                                                                           
+│       generate_samples: true                                                                                                                
+│       generated_samples_path: /home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/s
+│                                                                                                                                             
+├── optim
+│   └── weight_decay: 0                                                                                                                       
+│       lr: 0.0003                                                                                                                            
+│       beta1: 0.9                                                                                                                            
+│       beta2: 0.999                                                                                                                          
+│       eps: 1.0e-08                                                                                                                          
+│                                                                                                                                             
+├── trainer
+│   └── _target_: lightning.Trainer                                                                                                           
+│       accelerator: cuda                                                                                                                     
+│       num_nodes: 1                                                                                                                          
+│       devices: 1                                                                                                                            
+│       accumulate_grad_batches: 256                                                                                                          
+│       gradient_clip_val: 1.0                                                                                                                
+│       precision: bf16                                                                                                                       
+│       num_sanity_val_steps: 2                                                                                                               
+│       max_steps: 1000000                                                                                                                    
+│       log_every_n_steps: 100                                                                                                                
+│       limit_train_batches: 1.0                                                                                                              
+│       limit_val_batches: 1.0                                                                                                                
+│       val_check_interval: 5000                                                                                                              
+│       check_val_every_n_epoch: null                                                                                                         
+│                                                                                                                                             
+├── wandb
+│   └── project: duo                                                                                                                          
+│       group: null                                                                                                                           
+│       job_type: null                                                                                                                        
+│       name: null                                                                                                                            
+│       id: None_42                                                                                                                           
+│       tags:                                                                                                                                 
+│       - log-linear                                                                                                                          
+│       - lm1b                                                                                                                                
+│       - lm1b                                                                                                                                
+│       - duo_base                                                                                                                            
+│       offline: true                                                                                                                         
+│                                                                                                                                             
+├── checkpointing
+│   └── save_dir: /home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_1/upstrea
+│       resume_from_ckpt: true                                                                                                                
+│       resume_ckpt_path: /home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_1
+│                                                                                                                                             
+├── callbacks
+│   └── checkpoint_every_n_steps:                                                                                                             
+│         _target_: lightning.pytorch.callbacks.ModelCheckpoint                                                                               
+│         save_top_k: -1                                                                                                                      
+│         save_last: true                                                                                                                     
+│         dirpath: /home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_1/upstre
+│         verbose: true                                                                                                                       
+│         auto_insert_metric_name: false                                                                                                      
+│         every_n_train_steps: 500                                                                                                            
+│       checkpoint_monitor:                                                                                                                   
+│         _target_: lightning.pytorch.callbacks.ModelCheckpoint                                                                               
+│         monitor: val/nll                                                                                                                    
+│         mode: min                                                                                                                           
+│         save_top_k: 1                                                                                                                       
+│         save_last: false                                                                                                                    
+│         dirpath: /home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_1/upstre
+│         filename: best                                                                                                                      
+│         auto_insert_metric_name: false                                                                                                      
+│         verbose: true                                                                                                                       
+│       learning_rate_monitor:                                                                                                                
+│         _target_: lightning.pytorch.callbacks.LearningRateMonitor                                                                           
+│         logging_interval: step                                                                                                              
+│                                                                                                                                             
+├── data
+│   └── train: lm1b                                                                                                                           
+│       valid: lm1b                                                                                                                           
+│       modality: text                                                                                                                        
+│       tokenizer_name_or_path: bert-base-uncased                                                                                             
+│       cache_dir: /share/kuleshov/ssahoo/textdiffusion/data                                                                                  
+│       wrap: true                                                                                                                            
+│       streaming: false                                                                                                                      
+│       insert_train_eos: true                                                                                                                
+│       insert_valid_eos: true                                                                                                                
+│                                                                                                                                             
+├── model
+│   └── name: small                                                                                                                           
+│       type: ddit                                                                                                                            
+│       hidden_size: 768                                                                                                                      
+│       cond_dim: 128                                                                                                                         
+│       length: 128                                                                                                                           
+│       n_blocks: 12                                                                                                                          
+│       n_heads: 12                                                                                                                           
+│       scale_by_sigma: true                                                                                                                  
+│       dropout: 0.1                                                                                                                          
+│       tie_word_embeddings: false                                                                                                            
+│       vocab_lookup: true                                                                                                                    
+│                                                                                                                                             
+├── strategy
+│   └── _target_: lightning.pytorch.strategies.DDPStrategy                                                                                    
+│       find_unused_parameters: false                                                                                                         
+│                                                                                                                                             
+├── noise
+│   └── type: log-linear                                                                                                                      
+│       eps: 0.001                                                                                                                            
+│                                                                                                                                             
+├── lr_scheduler
+│   └── _target_: transformers.get_constant_schedule_with_warmup                                                                              
+│       num_warmup_steps: 2500                                                                                                                
+│                                                                                                                                             
+├── prior
+│   └── type: none                                                                                                                            
+│       latent_width: 0                                                                                                                       
+│       latent_height: 0                                                                                                                      
+│                                                                                                                                             
+└── algo
+    └── name: duo_base                                                                                                                        
+        backbone: dit                                                                                                                         
+        parameterization: mean                                                                                                                
+        time_conditioning: true                                                                                                               
+        T: 0                                                                                                                                  
+        subs_masking: false                                                                                                                   
+        causal_attention: false                                                                                                               
+        ignore_bos: false                                                                                                                     
+        loss_type: elbo                                                                                                                       
+                                                                                                                                              

results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_1/upstream_hydra/dlb_pinned_upstream_main.log

@@ -0,0 +1,4 @@
+[2026-08-25 10:38:05,643][dlb_pinned_upstream_main][INFO] - Starting Sample Eval.
+[2026-08-25 10:38:13,066][dlb_pinned_upstream_main][INFO] - Generative perplexity: 1125.733642578125
+[2026-08-25 10:38:13,066][dlb_pinned_upstream_main][INFO] - Sample entropy: 4.429431915283203
+[2026-08-25 10:38:13,067][dlb_pinned_upstream_main][INFO] - Samples saved at: /home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_1/upstream_samples.json

results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_2/upstream_hydra/.hydra/overrides.yaml

@@ -0,0 +1,19 @@
+- mode=sample_eval
+- seed=42
+- data=lm1b-wrap
+- model=small
+- model.length=128
+- loader.batch_size=2
+- loader.eval_batch_size=1
+- sampling.num_sample_batches=1
+- sampling.steps=2
+- eval.checkpoint_path=/home/ruizzhao/diffusion_baseline/checkpoints/reference_reproduction/duo_dcd/lm1b/model.ckpt
+- trainer.devices=1
+- checkpointing.save_dir=/home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_2/upstream_hydra
+- +wandb.offline=true
+- eval.generated_samples_path=/home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_2/upstream_samples.json
+- algo=duo_base
+- algo.backbone=dit
+- sampling.predictor=ancestral
+- sampling.noise_removal=ancestral
+- eval.disable_ema=False

results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_32/upstream_hydra/.hydra/overrides.yaml

@@ -0,0 +1,19 @@
+- mode=sample_eval
+- seed=42
+- data=lm1b-wrap
+- model=small
+- model.length=128
+- loader.batch_size=2
+- loader.eval_batch_size=1
+- sampling.num_sample_batches=1
+- sampling.steps=32
+- eval.checkpoint_path=/home/ruizzhao/diffusion_baseline/checkpoints/reference_reproduction/duo_dcd/lm1b/model.ckpt
+- trainer.devices=1
+- checkpointing.save_dir=/home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_32/upstream_hydra
+- +wandb.offline=true
+- eval.generated_samples_path=/home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_32/upstream_samples.json
+- algo=duo_base
+- algo.backbone=dit
+- sampling.predictor=ancestral
+- sampling.noise_removal=ancestral
+- eval.disable_ema=False

results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_32/upstream_hydra/dlb_pinned_upstream_main.log

@@ -0,0 +1,4 @@
+[2026-08-25 10:47:54,997][dlb_pinned_upstream_main][INFO] - Starting Sample Eval.
+[2026-08-25 10:48:08,607][dlb_pinned_upstream_main][INFO] - Generative perplexity: 94.98780822753906
+[2026-08-25 10:48:08,608][dlb_pinned_upstream_main][INFO] - Sample entropy: 4.304296016693115
+[2026-08-25 10:48:08,608][dlb_pinned_upstream_main][INFO] - Samples saved at: /home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_32/upstream_samples.json

results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_4/upstream_hydra/config_tree.txt

@@ -0,0 +1,169 @@
+CONFIG
+├── mode
+│   └── sample_eval                                                                                                                           
+├── seed
+│   └── 42                                                                                                                                    
+├── loader
+│   └── global_batch_size: 512                                                                                                                
+│       eval_global_batch_size: 512                                                                                                           
+│       batch_size: 2                                                                                                                         
+│       eval_batch_size: 1                                                                                                                    
+│       num_workers: 512                                                                                                                      
+│       pin_memory: true                                                                                                                      
+│                                                                                                                                             
+├── sampling
+│   └── predictor: ancestral                                                                                                                  
+│       steps: 4                                                                                                                              
+│       noise_removal: ancestral                                                                                                              
+│       use_float64: true                                                                                                                     
+│       p_nucleus: 1.0                                                                                                                        
+│       num_sample_batches: 1                                                                                                                 
+│       num_sample_log: 2                                                                                                                     
+│       semi_ar: false                                                                                                                        
+│       stride_length: 1                                                                                                                      
+│       num_strides: 1                                                                                                                        
+│       guid_weight: null                                                                                                                     
+│       psi:                                                                                                                                  
+│         time_profile: linear                                                                                                                
+│         high_mode: pure-posterior                                                                                                           
+│         middle_mode: pure-posterior                                                                                                         
+│         low_mode: pure-posterior                                                                                                            
+│         high_frac: 0.2                                                                                                                      
+│         middle_frac: 0.6                                                                                                                    
+│                                                                                                                                             
+├── training
+│   └── ema: 0.9999                                                                                                                           
+│       antithetic_sampling: true                                                                                                             
+│       importance_sampling: false                                                                                                            
+│       sampling_eps: 0.001                                                                                                                   
+│       change_of_variables: false                                                                                                            
+│       loss_precision: bf16                                                                                                                  
+│       finetune_path: ''                                                                                                                     
+│       class_dropout_p: 0.1                                                                                                                  
+│                                                                                                                                             
+├── eval
+│   └── checkpoint_path: /home/ruizzhao/diffusion_baseline/checkpoints/reference_reproduction/duo_dcd/lm1b/model.ckpt                         
+│       disable_ema: false                                                                                                                    
+│       compute_generative_perplexity: false                                                                                                  
+│       perplexity_batch_size: 8                                                                                                              
+│       compute_perplexity_on_sanity: false                                                                                                   
+│       gen_ppl_eval_model_name_or_path: gpt2-large                                                                                           
+│       generate_samples: true                                                                                                                
+│       generated_samples_path: /home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/s
+│                                                                                                                                             
+├── optim
+│   └── weight_decay: 0                                                                                                                       
+│       lr: 0.0003                                                                                                                            
+│       beta1: 0.9                                                                                                                            
+│       beta2: 0.999                                                                                                                          
+│       eps: 1.0e-08                                                                                                                          
+│                                                                                                                                             
+├── trainer
+│   └── _target_: lightning.Trainer                                                                                                           
+│       accelerator: cuda                                                                                                                     
+│       num_nodes: 1                                                                                                                          
+│       devices: 1                                                                                                                            
+│       accumulate_grad_batches: 256                                                                                                          
+│       gradient_clip_val: 1.0                                                                                                                
+│       precision: bf16                                                                                                                       
+│       num_sanity_val_steps: 2                                                                                                               
+│       max_steps: 1000000                                                                                                                    
+│       log_every_n_steps: 100                                                                                                                
+│       limit_train_batches: 1.0                                                                                                              
+│       limit_val_batches: 1.0                                                                                                                
+│       val_check_interval: 5000                                                                                                              
+│       check_val_every_n_epoch: null                                                                                                         
+│                                                                                                                                             
+├── wandb
+│   └── project: duo                                                                                                                          
+│       group: null                                                                                                                           
+│       job_type: null                                                                                                                        
+│       name: null                                                                                                                            
+│       id: None_42                                                                                                                           
+│       tags:                                                                                                                                 
+│       - log-linear                                                                                                                          
+│       - lm1b                                                                                                                                
+│       - lm1b                                                                                                                                
+│       - duo_base                                                                                                                            
+│       offline: true                                                                                                                         
+│                                                                                                                                             
+├── checkpointing
+│   └── save_dir: /home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_4/upstrea
+│       resume_from_ckpt: true                                                                                                                
+│       resume_ckpt_path: /home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_4
+│                                                                                                                                             
+├── callbacks
+│   └── checkpoint_every_n_steps:                                                                                                             
+│         _target_: lightning.pytorch.callbacks.ModelCheckpoint                                                                               
+│         save_top_k: -1                                                                                                                      
+│         save_last: true                                                                                                                     
+│         dirpath: /home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_4/upstre
+│         verbose: true                                                                                                                       
+│         auto_insert_metric_name: false                                                                                                      
+│         every_n_train_steps: 500                                                                                                            
+│       checkpoint_monitor:                                                                                                                   
+│         _target_: lightning.pytorch.callbacks.ModelCheckpoint                                                                               
+│         monitor: val/nll                                                                                                                    
+│         mode: min                                                                                                                           
+│         save_top_k: 1                                                                                                                       
+│         save_last: false                                                                                                                    
+│         dirpath: /home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_4/upstre
+│         filename: best                                                                                                                      
+│         auto_insert_metric_name: false                                                                                                      
+│         verbose: true                                                                                                                       
+│       learning_rate_monitor:                                                                                                                
+│         _target_: lightning.pytorch.callbacks.LearningRateMonitor                                                                           
+│         logging_interval: step                                                                                                              
+│                                                                                                                                             
+├── data
+│   └── train: lm1b                                                                                                                           
+│       valid: lm1b                                                                                                                           
+│       modality: text                                                                                                                        
+│       tokenizer_name_or_path: bert-base-uncased                                                                                             
+│       cache_dir: /share/kuleshov/ssahoo/textdiffusion/data                                                                                  
+│       wrap: true                                                                                                                            
+│       streaming: false                                                                                                                      
+│       insert_train_eos: true                                                                                                                
+│       insert_valid_eos: true                                                                                                                
+│                                                                                                                                             
+├── model
+│   └── name: small                                                                                                                           
+│       type: ddit                                                                                                                            
+│       hidden_size: 768                                                                                                                      
+│       cond_dim: 128                                                                                                                         
+│       length: 128                                                                                                                           
+│       n_blocks: 12                                                                                                                          
+│       n_heads: 12                                                                                                                           
+│       scale_by_sigma: true                                                                                                                  
+│       dropout: 0.1                                                                                                                          
+│       tie_word_embeddings: false                                                                                                            
+│       vocab_lookup: true                                                                                                                    
+│                                                                                                                                             
+├── strategy
+│   └── _target_: lightning.pytorch.strategies.DDPStrategy                                                                                    
+│       find_unused_parameters: false                                                                                                         
+│                                                                                                                                             
+├── noise
+│   └── type: log-linear                                                                                                                      
+│       eps: 0.001                                                                                                                            
+│                                                                                                                                             
+├── lr_scheduler
+│   └── _target_: transformers.get_constant_schedule_with_warmup                                                                              
+│       num_warmup_steps: 2500                                                                                                                
+│                                                                                                                                             
+├── prior
+│   └── type: none                                                                                                                            
+│       latent_width: 0                                                                                                                       
+│       latent_height: 0                                                                                                                      
+│                                                                                                                                             
+└── algo
+    └── name: duo_base                                                                                                                        
+        backbone: dit                                                                                                                         
+        parameterization: mean                                                                                                                
+        time_conditioning: true                                                                                                               
+        T: 0                                                                                                                                  
+        subs_masking: false                                                                                                                   
+        causal_attention: false                                                                                                               
+        ignore_bos: false                                                                                                                     
+        loss_type: elbo                                                                                                                       
+                                                                                                                                              

results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_8/upstream_hydra/config_tree.txt

@@ -0,0 +1,169 @@
+CONFIG
+├── mode
+│   └── sample_eval                                                                                                                           
+├── seed
+│   └── 42                                                                                                                                    
+├── loader
+│   └── global_batch_size: 512                                                                                                                
+│       eval_global_batch_size: 512                                                                                                           
+│       batch_size: 2                                                                                                                         
+│       eval_batch_size: 1                                                                                                                    
+│       num_workers: 512                                                                                                                      
+│       pin_memory: true                                                                                                                      
+│                                                                                                                                             
+├── sampling
+│   └── predictor: ancestral                                                                                                                  
+│       steps: 8                                                                                                                              
+│       noise_removal: ancestral                                                                                                              
+│       use_float64: true                                                                                                                     
+│       p_nucleus: 1.0                                                                                                                        
+│       num_sample_batches: 1                                                                                                                 
+│       num_sample_log: 2                                                                                                                     
+│       semi_ar: false                                                                                                                        
+│       stride_length: 1                                                                                                                      
+│       num_strides: 1                                                                                                                        
+│       guid_weight: null                                                                                                                     
+│       psi:                                                                                                                                  
+│         time_profile: linear                                                                                                                
+│         high_mode: pure-posterior                                                                                                           
+│         middle_mode: pure-posterior                                                                                                         
+│         low_mode: pure-posterior                                                                                                            
+│         high_frac: 0.2                                                                                                                      
+│         middle_frac: 0.6                                                                                                                    
+│                                                                                                                                             
+├── training
+│   └── ema: 0.9999                                                                                                                           
+│       antithetic_sampling: true                                                                                                             
+│       importance_sampling: false                                                                                                            
+│       sampling_eps: 0.001                                                                                                                   
+│       change_of_variables: false                                                                                                            
+│       loss_precision: bf16                                                                                                                  
+│       finetune_path: ''                                                                                                                     
+│       class_dropout_p: 0.1                                                                                                                  
+│                                                                                                                                             
+├── eval
+│   └── checkpoint_path: /home/ruizzhao/diffusion_baseline/checkpoints/reference_reproduction/duo_dcd/lm1b/model.ckpt                         
+│       disable_ema: false                                                                                                                    
+│       compute_generative_perplexity: false                                                                                                  
+│       perplexity_batch_size: 8                                                                                                              
+│       compute_perplexity_on_sanity: false                                                                                                   
+│       gen_ppl_eval_model_name_or_path: gpt2-large                                                                                           
+│       generate_samples: true                                                                                                                
+│       generated_samples_path: /home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/s
+│                                                                                                                                             
+├── optim
+│   └── weight_decay: 0                                                                                                                       
+│       lr: 0.0003                                                                                                                            
+│       beta1: 0.9                                                                                                                            
+│       beta2: 0.999                                                                                                                          
+│       eps: 1.0e-08                                                                                                                          
+│                                                                                                                                             
+├── trainer
+│   └── _target_: lightning.Trainer                                                                                                           
+│       accelerator: cuda                                                                                                                     
+│       num_nodes: 1                                                                                                                          
+│       devices: 1                                                                                                                            
+│       accumulate_grad_batches: 256                                                                                                          
+│       gradient_clip_val: 1.0                                                                                                                
+│       precision: bf16                                                                                                                       
+│       num_sanity_val_steps: 2                                                                                                               
+│       max_steps: 1000000                                                                                                                    
+│       log_every_n_steps: 100                                                                                                                
+│       limit_train_batches: 1.0                                                                                                              
+│       limit_val_batches: 1.0                                                                                                                
+│       val_check_interval: 5000                                                                                                              
+│       check_val_every_n_epoch: null                                                                                                         
+│                                                                                                                                             
+├── wandb
+│   └── project: duo                                                                                                                          
+│       group: null                                                                                                                           
+│       job_type: null                                                                                                                        
+│       name: null                                                                                                                            
+│       id: None_42                                                                                                                           
+│       tags:                                                                                                                                 
+│       - log-linear                                                                                                                          
+│       - lm1b                                                                                                                                
+│       - lm1b                                                                                                                                
+│       - duo_base                                                                                                                            
+│       offline: true                                                                                                                         
+│                                                                                                                                             
+├── checkpointing
+│   └── save_dir: /home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_8/upstrea
+│       resume_from_ckpt: true                                                                                                                
+│       resume_ckpt_path: /home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_8
+│                                                                                                                                             
+├── callbacks
+│   └── checkpoint_every_n_steps:                                                                                                             
+│         _target_: lightning.pytorch.callbacks.ModelCheckpoint                                                                               
+│         save_top_k: -1                                                                                                                      
+│         save_last: true                                                                                                                     
+│         dirpath: /home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_8/upstre
+│         verbose: true                                                                                                                       
+│         auto_insert_metric_name: false                                                                                                      
+│         every_n_train_steps: 500                                                                                                            
+│       checkpoint_monitor:                                                                                                                   
+│         _target_: lightning.pytorch.callbacks.ModelCheckpoint                                                                               
+│         monitor: val/nll                                                                                                                    
+│         mode: min                                                                                                                           
+│         save_top_k: 1                                                                                                                       
+│         save_last: false                                                                                                                    
+│         dirpath: /home/ruizzhao/diffusion_baseline/results_duo_dcd_ckpt_probe/best_seed42/timing_scratch/samples/lm1b/duo_dcd/steps_8/upstre
+│         filename: best                                                                                                                      
+│         auto_insert_metric_name: false                                                                                                      
+│         verbose: true                                                                                                                       
+│       learning_rate_monitor:                                                                                                                
+│         _target_: lightning.pytorch.callbacks.LearningRateMonitor                                                                           
+│         logging_interval: step                                                                                                              
+│                                                                                                                                             
+├── data
+│   └── train: lm1b                                                                                                                           
+│       valid: lm1b                                                                                                                           
+│       modality: text                                                                                                                        
+│       tokenizer_name_or_path: bert-base-uncased                                                                                             
+│       cache_dir: /share/kuleshov/ssahoo/textdiffusion/data                                                                                  
+│       wrap: true                                                                                                                            
+│       streaming: false                                                                                                                      
+│       insert_train_eos: true                                                                                                                
+│       insert_valid_eos: true                                                                                                                
+│                                                                                                                                             
+├── model
+│   └── name: small                                                                                                                           
+│       type: ddit                                                                                                                            
+│       hidden_size: 768                                                                                                                      
+│       cond_dim: 128                                                                                                                         
+│       length: 128                                                                                                                           
+│       n_blocks: 12                                                                                                                          
+│       n_heads: 12                                                                                                                           
+│       scale_by_sigma: true                                                                                                                  
+│       dropout: 0.1                                                                                                                          
+│       tie_word_embeddings: false                                                                                                            
+│       vocab_lookup: true                                                                                                                    
+│                                                                                                                                             
+├── strategy
+│   └── _target_: lightning.pytorch.strategies.DDPStrategy                                                                                    
+│       find_unused_parameters: false                                                                                                         
+│                                                                                                                                             
+├── noise
+│   └── type: log-linear                                                                                                                      
+│       eps: 0.001                                                                                                                            
+│                                                                                                                                             
+├── lr_scheduler
+│   └── _target_: transformers.get_constant_schedule_with_warmup                                                                              
+│       num_warmup_steps: 2500                                                                                                                
+│                                                                                                                                             
+├── prior
+│   └── type: none                                                                                                                            
+│       latent_width: 0                                                                                                                       
+│       latent_height: 0                                                                                                                      
+│                                                                                                                                             
+└── algo
+    └── name: duo_base                                                                                                                        
+        backbone: dit                                                                                                                         
+        parameterization: mean                                                                                                                
+        time_conditioning: true                                                                                                               
+        T: 0                                                                                                                                  
+        subs_masking: false                                                                                                                   
+        causal_attention: false                                                                                                               
+        ignore_bos: false                                                                                                                     
+        loss_type: elbo                                                                                                                       
+                                                                                                                                              

scripts/aggregate_results.py

@@ -0,0 +1,43 @@
+#!/usr/bin/env python
+"""Publish strict or partial baseline summary tables."""
+
+from __future__ import annotations
+
+import argparse
+from pathlib import Path
+
+from dlb.aggregate import IncompleteMatrixError, aggregate
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--root", type=Path, default=Path.cwd())
+    parser.add_argument(
+        "--output-dir", type=Path, default=None,
+        help="defaults to ROOT/results/summary",
+    )
+    parser.add_argument("--partial", action="store_true")
+    args = parser.parse_args(argv)
+    root = args.root.resolve()
+    output = (args.output_dir or root / "results/summary").resolve()
+    try:
+        report = aggregate(
+            root,
+            strict=not args.partial,
+            partial=args.partial,
+            output_dir=output,
+        )
+    except IncompleteMatrixError as error:
+        print(str(error))
+        print(f"failures={len(error.failures)}")
+        return 1
+    print(
+        f"complete={str(report.complete).lower()} "
+        f"rows={len(report.rows)} failures={len(report.failures)} "
+        f"unsupported={len(report.unsupported)} output={output}"
+    )
+    return 0 if report.complete else 1
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

scripts/audit_elf_inputs.py

@@ -0,0 +1,115 @@
+#!/usr/bin/env python3
+"""Offline CPU audit of raw versus released ELF conditional validation inputs."""
+from __future__ import annotations
+
+import argparse
+from collections import Counter, defaultdict
+import json
+from pathlib import Path
+
+from elf_common import (ROOT, asset, condition_token_ids, offline, read_jsonl,
+                        require_server, sha256, write_json)
+
+
+def audit_rows(raw, official, tokenizer, limit, task):
+    def inspect(rows):
+        token_rows = [condition_token_ids(row, tokenizer) for row in rows]
+        return token_rows, {
+            "rows": len(rows),
+            "blank_text_ids": [row["id"] for row in rows if not row["input"].strip()],
+            "empty_condition_ids": [row["id"] for row, ids in zip(rows, token_rows) if not ids],
+            "truncated_sources": sum(len(ids) > limit for ids in token_rows),
+            "condition_length_mean_before_cap": sum(map(len, token_rows)) / len(rows) if rows else None,
+        }
+
+    raw_ids, raw_stats = inspect(raw)
+    official_ids, official_stats = inspect(official)
+    by_pair = defaultdict(list)
+    for row, ids in zip(official, official_ids):
+        by_pair[(row["input"], row["output"])].append((row, ids))
+    matches = Counter()
+    examples = []
+    for row, ids in zip(raw, raw_ids):
+        candidates = by_pair.get((row["input"], row["output"]), [])
+        if len(candidates) != 1:
+            matches["unmatched_text_pair" if not candidates else "ambiguous_text_pair"] += 1
+            continue
+        author, expected = candidates[0]
+        matches["unique_text_pair"] += 1
+        matches["exact_ids"] += int(ids == expected)
+        matches["exact_ids_after_source_cap"] += int(ids[:limit] == expected[:limit])
+        matches["raw_plus_eos_equals_official"] += int(ids + [tokenizer.eos_token_id] == expected)
+        if ids[:limit] != expected[:limit] and len(examples) < 5:
+            examples.append({
+                "raw_id": row["id"], "official_id": author["id"],
+                "source_excerpt": row["input"][:160],
+                "raw_length": len(ids), "official_length": len(expected),
+                "raw_first_ids": ids[:16], "raw_last_ids": ids[-16:],
+                "official_first_ids": expected[:16], "official_last_ids": expected[-16:],
+                "official_condition_decoded_excerpt": tokenizer.decode(expected[:limit], skip_special_tokens=False)[:240],
+            })
+    corrected = [condition_token_ids(row, tokenizer, task=task) for row in raw]
+    # Multisets account for every duplicate's multiplicity instead of skipping
+    # the eight WMT rows that the unique-pair diagnostic cannot disambiguate.
+    def records(rows, token_rows, capped):
+        return [(r["input"], r["output"], tuple(ids[:limit] if capped else ids))
+                for r, ids in zip(rows, token_rows)]
+    corrected_records = records(raw, corrected, False)
+    author_records = records(official, official_ids, False)
+    corrected_capped = records(raw, corrected, True)
+    author_capped = records(official, official_ids, True)
+    coverage = {
+        "raw_rows": len(raw), "official_rows": len(official),
+        "full_token_multisets_equal": Counter(corrected_records) == Counter(author_records),
+        "capped_token_multisets_equal": Counter(corrected_capped) == Counter(author_capped),
+        "same_order_and_full_tokens": corrected_records == author_records,
+        "same_order_and_capped_tokens": corrected_capped == author_capped,
+        "empty_condition_ids": [r["id"] for r, ids in zip(raw, corrected) if not ids],
+        "policy": "released_ids_or_raw_plus_eos_then_source_cap_v2",
+    }
+    coverage["ready_for_raw_validation"] = (
+        bool(raw) and coverage["capped_token_multisets_equal"] and not coverage["empty_condition_ids"])
+    return {"source_cap": limit, "raw": raw_stats, "official": official_stats,
+            "comparison_by_exact_source_and_reference_text_not_row_number": dict(matches),
+            "mismatch_examples": examples,
+            "official_conditions_nonempty": bool(official) and not official_stats["empty_condition_ids"],
+            "corrected_runner_comparison_including_duplicates": coverage}
+
+
+def load_checked(path, task):
+    meta = json.loads(path.with_suffix(".manifest.json").read_text())
+    rows = read_jsonl(path)
+    if (meta["sha256"] != sha256(path) or meta["task"] != task or meta["count"] != len(rows)
+            or len({row["id"] for row in rows}) != len(rows)):
+        raise ValueError(f"Input manifest/count/IDs mismatch: {path}")
+    return rows, {"path": str(path), "manifest": meta}
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--root", type=Path, default=ROOT)
+    parser.add_argument("--output", type=Path, required=True)
+    args = parser.parse_args()
+    require_server()
+    offline()
+    from transformers import AutoTokenizer
+    root = args.root.resolve()
+    tokenizer = AutoTokenizer.from_pretrained(str(asset(root, "t5")), local_files_only=True)
+    report = {"schema": "elf-condition-audit-v2", "tasks": {},
+              "auditor_sha256": sha256(Path(__file__)),
+              "helper_sha256": sha256(Path(__file__).with_name("elf_common.py")),
+              "assets_manifest_sha256": sha256(root / "data/elf/assets.json")}
+    for task, limit in (("wmt14", 64), ("xsum", 1024)):
+        raw, raw_binding = load_checked(root / f"data/elf/{task}-validation.jsonl", task)
+        official, official_binding = load_checked(root / f"data/elf/{task}-official-validation.jsonl", task)
+        print(f"Auditing {task}: raw={len(raw)}, official={len(official)}; no GPU inference", flush=True)
+        result = audit_rows(raw, official, tokenizer, limit, task)
+        result["inputs"] = {"raw": raw_binding, "official": official_binding}
+        report["tasks"][task] = result
+        print(json.dumps({k: v for k, v in result.items() if k != "inputs"}, indent=2, ensure_ascii=False), flush=True)
+    write_json(args.output, report)
+    print(f"Saved {args.output}. Input files unchanged; compared the corrected runner's EOS policy without generating samples.")
+
+
+if __name__ == "__main__":
+    main()

scripts/benchmark_4gpu.sh

@@ -0,0 +1,48 @@
+#!/usr/bin/env bash
+# Four-GPU local timing pass over the canonical generation matrix.
+set -uo pipefail
+
+SCRIPT_DIR="$(CDPATH= cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd -P)"
+DLB_ROOT="${DLB_ROOT:-$(CDPATH= cd -- "$SCRIPT_DIR/.." && pwd -P)}"
+PYTHON_BIN="${DLB_PYTHON:-python}"
+MATRIX="${DLB_MATRIX:-}"
+CUSTOM_MATRIX=0
+if [[ -n "$MATRIX" ]]; then CUSTOM_MATRIX=1; fi
+SEED=42
+GPUS="${DLB_GPUS:-0,1,2,3}"
+MAX_JOBS="${DLB_MAX_JOBS:-4}"
+DRY_RUN=()
+
+usage() {
+  printf 'Usage: %s [--root PATH] [--matrix PATH] [--seed N] [--gpus LIST] [--max-jobs N] [--dry-run]\n' "$0"
+}
+
+while (($#)); do
+  case "$1" in
+    --root) DLB_ROOT="$2"; shift 2 ;;
+    --matrix) MATRIX="$2"; CUSTOM_MATRIX=1; shift 2 ;;
+    --seed) SEED="$2"; shift 2 ;;
+    --gpus) GPUS="$2"; shift 2 ;;
+    --max-jobs) MAX_JOBS="$2"; shift 2 ;;
+    --dry-run) DRY_RUN=(--dry-run); shift ;;
+    --help) usage; exit 0 ;;
+    *) echo "unknown argument: $1" >&2; usage; exit 2 ;;
+  esac
+done
+
+DLB_ROOT="$(CDPATH= cd -- "$DLB_ROOT" && pwd -P)"
+if [[ -z "$MATRIX" ]]; then MATRIX="$DLB_ROOT/results/matrix/generation.tsv"; fi
+export DLB_ROOT
+mkdir -p "$DLB_ROOT/results/matrix" "$DLB_ROOT/results/logs"
+if (( CUSTOM_MATRIX )); then
+  if [[ ! -f "$MATRIX" ]]; then
+    echo "custom matrix does not exist: $MATRIX" >&2
+    exit 2
+  fi
+else
+  PYTHONPATH="$DLB_ROOT/src${PYTHONPATH:+:$PYTHONPATH}" "$PYTHON_BIN" -m dlb.matrix \
+    --root "$DLB_ROOT" --output "$MATRIX" --sample-count 1024 --seed "$SEED" || exit $?
+fi
+exec env PYTHONPATH="$DLB_ROOT/src${PYTHONPATH:+:$PYTHONPATH}" "$PYTHON_BIN" -m dlb.gpu_matrix \
+  --root "$DLB_ROOT" --stage benchmark --matrix "$MATRIX" \
+  --gpus "$GPUS" --max-jobs "$MAX_JOBS" --precision author "${DRY_RUN[@]}"

scripts/benchmark_conditional_4gpu.sh

@@ -0,0 +1,48 @@
+#!/usr/bin/env bash
+# Four-GPU local timing pass over the C64 conditional matrix.
+set -uo pipefail
+
+SCRIPT_DIR="$(CDPATH= cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd -P)"
+DLB_ROOT="${DLB_ROOT:-$(CDPATH= cd -- "$SCRIPT_DIR/.." && pwd -P)}"
+PYTHON_BIN="${DLB_PYTHON:-python}"
+MATRIX="${DLB_CONDITIONAL_MATRIX:-}"
+CUSTOM_MATRIX=0
+if [[ -n "$MATRIX" ]]; then CUSTOM_MATRIX=1; fi
+SEED=42
+GPUS="${DLB_GPUS:-0,1,2,3}"
+MAX_JOBS="${DLB_MAX_JOBS:-4}"
+DRY_RUN=()
+
+usage() {
+  printf 'Usage: %s [--root PATH] [--matrix PATH] [--seed N] [--gpus LIST] [--max-jobs N] [--dry-run]\n' "$0"
+}
+
+while (($#)); do
+  case "$1" in
+    --root) DLB_ROOT="$2"; shift 2 ;;
+    --matrix) MATRIX="$2"; CUSTOM_MATRIX=1; shift 2 ;;
+    --seed) SEED="$2"; shift 2 ;;
+    --gpus) GPUS="$2"; shift 2 ;;
+    --max-jobs) MAX_JOBS="$2"; shift 2 ;;
+    --dry-run) DRY_RUN=(--dry-run); shift ;;
+    --help) usage; exit 0 ;;
+    *) echo "unknown argument: $1" >&2; usage; exit 2 ;;
+  esac
+done
+
+DLB_ROOT="$(CDPATH= cd -- "$DLB_ROOT" && pwd -P)"
+if [[ -z "$MATRIX" ]]; then MATRIX="$DLB_ROOT/results/conditional/matrix/generation.tsv"; fi
+export DLB_ROOT
+mkdir -p "$DLB_ROOT/results/conditional/matrix" "$DLB_ROOT/results/logs"
+if (( CUSTOM_MATRIX )); then
+  if [[ ! -f "$MATRIX" ]]; then
+    echo "custom matrix does not exist: $MATRIX" >&2
+    exit 2
+  fi
+else
+  PYTHONPATH="$DLB_ROOT/src${PYTHONPATH:+:$PYTHONPATH}" "$PYTHON_BIN" -m dlb.conditional_matrix \
+    --root "$DLB_ROOT" --output "$MATRIX" || exit $?
+fi
+exec env PYTHONPATH="$DLB_ROOT/src${PYTHONPATH:+:$PYTHONPATH}" "$PYTHON_BIN" -m dlb.gpu_matrix \
+  --root "$DLB_ROOT" --stage benchmark --matrix "$MATRIX" \
+  --gpus "$GPUS" --max-jobs "$MAX_JOBS" --precision author "${DRY_RUN[@]}"

scripts/benchmark_conditional_all.sh

@@ -0,0 +1,46 @@
+#!/usr/bin/env bash
+# Server-side serial timing pass over the C64 conditional matrix.
+set -uo pipefail
+
+SCRIPT_DIR="$(CDPATH= cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd -P)"
+DLB_ROOT="${DLB_ROOT:-$(CDPATH= cd -- "$SCRIPT_DIR/.." && pwd -P)}"
+PYTHON_BIN="${DLB_PYTHON:-python}"
+MATRIX="${DLB_CONDITIONAL_MATRIX:-}"
+SEED=42
+
+while (($#)); do
+  case "$1" in
+    --root) DLB_ROOT="$2"; shift 2 ;;
+    --matrix) MATRIX="$2"; shift 2 ;;
+    --seed) SEED="$2"; shift 2 ;;
+    --help)
+      printf 'Usage: %s [--root PATH] [--matrix PATH] [--seed N]\n' "$0"
+      exit 0
+      ;;
+    *) echo "unknown argument: $1" >&2; exit 2 ;;
+  esac
+done
+
+DLB_ROOT="$(CDPATH= cd -- "$DLB_ROOT" && pwd -P)"
+if [[ -z "$MATRIX" ]]; then MATRIX="$DLB_ROOT/results/conditional/matrix/generation.tsv"; fi
+export DLB_ROOT
+mkdir -p "$DLB_ROOT/results/conditional/matrix" "$DLB_ROOT/results/logs"
+PYTHONPATH="$DLB_ROOT/src${PYTHONPATH:+:$PYTHONPATH}" "$PYTHON_BIN" -m dlb.conditional_matrix \
+  --root "$DLB_ROOT" --output "$MATRIX" || exit $?
+
+failures="$DLB_ROOT/results/logs/conditional_benchmark_failures.tsv"
+printf 'task_id\tmodel\tdataset\tsteps\texit_code\n' >"$failures"
+status=0
+while IFS=$'\t' read -r task_id category model dataset steps sample_count seed environment adapter source provenance protocol conditioning_manifest conditioning_manifest_sha256 sample_dir metrics_path timing_path; do
+  [[ "$task_id" == "task_id" || -z "$task_id" ]] && continue
+  echo "CONDITIONAL BENCHMARK $task_id"
+  "$SCRIPT_DIR/benchmark_conditional_one.sh" --model "$model" --dataset "$dataset" \
+    --steps "$steps" --seed "$SEED" --precision author
+  exit_code=$?
+  if ((exit_code != 0)); then
+    printf '%s\t%s\t%s\t%s\t%s\n' "$task_id" "$model" "$dataset" "$steps" "$exit_code" >>"$failures"
+    status=1
+  fi
+done < <(tail -n +3 "$MATRIX")
+if [[ ! -s "$failures" ]]; then rm -f "$failures"; fi
+exit "$status"

scripts/benchmark_one.sh

@@ -0,0 +1,66 @@
+#!/usr/bin/env bash
+# Run one server-side, registry-bound primary-latency benchmark.
+set -u
+
+script_dir="$(CDPATH= cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd -P)"
+DLB_ROOT="${DLB_ROOT:-$(CDPATH= cd -- "$script_dir/.." && pwd -P)}"
+
+model=""
+dataset=""
+steps=""
+seed=""
+precision=""
+dry_run="false"
+
+usage() {
+  echo "usage: $0 --model MODEL --dataset {lm1b,owt} --steps N --seed N --precision author [--dry-run]" >&2
+}
+
+while (( $# )); do
+  case "$1" in
+    --model|--dataset|--steps|--seed|--precision)
+      if (( $# < 2 )); then usage; exit 2; fi
+      case "$1" in
+        --model) model="$2" ;;
+        --dataset) dataset="$2" ;;
+        --steps) steps="$2" ;;
+        --seed) seed="$2" ;;
+        --precision) precision="$2" ;;
+      esac
+      shift 2 ;;
+    --dry-run) dry_run="true"; shift ;;
+    --help) usage; exit 0 ;;
+    *) echo "unknown argument: $1" >&2; usage; exit 2 ;;
+  esac
+done
+
+if [[ -z "$model" || -z "$dataset" || -z "$steps" || -z "$seed" || -z "$precision" ]]; then
+  usage; exit 2
+fi
+if [[ ! "$steps" =~ ^[1-9][0-9]*$ || ! "$seed" =~ ^-?[0-9]+$ ]]; then
+  echo "steps must be positive and seed must be an integer" >&2; exit 2
+fi
+if [[ "$precision" != "author" ]]; then
+  echo "precision must be author (the pinned upstream inference policy)" >&2; exit 2
+fi
+
+DLB_PYTHON="${DLB_PYTHON:-python}"
+module_args=(--root "$DLB_ROOT" --models "$model" --datasets "$dataset" --steps "$steps" --seed "$seed" --precision "$precision")
+if [[ "$dry_run" == "true" ]]; then
+  exec env PYTHONPATH="$DLB_ROOT/src${PYTHONPATH:+:$PYTHONPATH}" "$DLB_PYTHON" -m dlb.benchmarking "${module_args[@]}" --dry-run
+fi
+
+if environment="$(PYTHONPATH="$DLB_ROOT/src${PYTHONPATH:+:$PYTHONPATH}" "$DLB_PYTHON" -m dlb.runner \
+  --root "$DLB_ROOT" --model "$model" --dataset "$dataset" --steps "$steps" \
+  --num-samples 1 --seed "$seed" --validate-only)"; then
+  :
+else
+  exit $?
+fi
+if [[ ! "$environment" =~ ^[a-z][a-z0-9-]*$ ]]; then
+  echo "registry validation returned an unsafe environment name" >&2; exit 2
+fi
+
+exec "${DLB_CONDA:-conda}" run -n "$environment" env \
+  PYTHONPATH="$DLB_ROOT/src${PYTHONPATH:+:$PYTHONPATH}" \
+  python -m dlb.benchmarking "${module_args[@]}"

scripts/distill/di4c.sh

@@ -0,0 +1,8 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd -P)"
+DLB_ROOT="$(cd -- "$SCRIPT_DIR/../.." && pwd -P)"
+export PYTHONPATH="$DLB_ROOT/src${PYTHONPATH:+:$PYTHONPATH}"
+PYTHON_BIN="${DLB_PYTHON:-python}"
+exec "$PYTHON_BIN" -m dlb.recipes --root "$DLB_ROOT" --recipe di4c "$@"

scripts/distill/fmlm.sh

@@ -0,0 +1,8 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd -P)"
+DLB_ROOT="$(cd -- "$SCRIPT_DIR/../.." && pwd -P)"
+export PYTHONPATH="$DLB_ROOT/src${PYTHONPATH:+:$PYTHONPATH}"
+PYTHON_BIN="${DLB_PYTHON:-python}"
+exec "$PYTHON_BIN" -m dlb.recipes --root "$DLB_ROOT" --recipe fmlm "$@"

scripts/elf_common.py

@@ -0,0 +1,131 @@
+"""Small, independent ELF experiment helpers (no GPU imports)."""
+from __future__ import annotations
+
+import hashlib
+import json
+import os
+from pathlib import Path
+import subprocess
+import sys
+
+ROOT = Path(__file__).resolve().parents[1]
+SOURCE_COMMIT = "b29d8833609e9ab7f67cd9da39435ac5cea04837"
+
+
+def sha256(path):
+    digest = hashlib.sha256()
+    with Path(path).open("rb") as handle:
+        for chunk in iter(lambda: handle.read(1024 * 1024), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def write_json(path, data):
+    path = Path(path)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    temporary = path.with_suffix(path.suffix + ".tmp")
+    temporary.write_text(json.dumps(data, indent=2, ensure_ascii=False, allow_nan=False) + "\n")
+    temporary.replace(path)
+
+
+def read_jsonl(path):
+    with Path(path).open() as handle:
+        return [json.loads(line) for line in handle if line.strip()]
+
+
+def condition_token_ids(row, tokenizer, *, task=None):
+    """Preserve released IDs; task-trained raw sources include the released EOS.
+
+    task=None retains plain tokenization for auditing the old input path.
+    Append before the model's source cap, exactly as the released Arrow inputs.
+    A C64 continuation is not an ended source sentence and keeps its old policy.
+    """
+    ids = row.get("condition_input_ids")
+    if ids is None:
+        ids = tokenizer(row["input"], add_special_tokens=False, verbose=False)["input_ids"]
+        if task in ("wmt14", "xsum"):
+            eos = tokenizer.eos_token_id
+            if type(eos) is not int or eos < 0:
+                raise ValueError("Task-trained ELF conditions require a valid EOS token")
+            ids = ids + [eos]
+    vocab_size = len(tokenizer)
+    if not isinstance(ids, list) or any(type(t) is not int or t < 0 or t >= vocab_size for t in ids):
+        raise ValueError(f"Invalid condition IDs at source_id={row.get('id')}")
+    return ids
+
+
+def preflight_conditions(rows, tokenizer, task, max_input_length, max_length):
+    """Validate the whole requested input before loading weights; never drop rows."""
+    prepared, failures = [], []
+    for index, row in enumerate(rows):
+        ids = condition_token_ids(row, tokenizer, task=task)
+        if task == "owt-prefix":
+            fits = 0 < len(ids) < max_length - 64
+        else:
+            ids = ids[:max_input_length]
+            fits = bool(ids)
+        if not fits:
+            failures.append({"row_index": index, "source_id": row.get("id"), "tokens": len(ids)})
+        prepared.append(ids)
+    if failures:
+        raise ValueError(
+            f"Invalid/empty ELF conditions: {len(failures)}/{len(rows)}; "
+            f"first affected rows (zero-based): {failures[:20]}. No samples were dropped. "
+            "Run scripts/audit_elf_inputs.py to compare raw and official inputs; "
+            "do not invent source text or delete rows to bypass this check."
+        )
+    return prepared
+
+
+def require_server():
+    if sys.platform != "linux":
+        raise RuntimeError("ELF downloads and GPU experiments must run on the Linux server.")
+
+
+def source_path(root):
+    path = Path(root) / "upstreams/elf"
+    commit = subprocess.check_output(["git", "-C", str(path), "rev-parse", "HEAD"], text=True).strip()
+    dirty = subprocess.check_output(["git", "-C", str(path), "status", "--porcelain"], text=True)
+    if commit != SOURCE_COMMIT or dirty.strip():
+        raise ValueError("ELF upstream must be clean and match the locked commit")
+    return path
+
+
+def default_sampling(task, steps):
+    if steps < 1:
+        raise ValueError("steps must be positive")
+    if task in ("wmt14", "xsum"):
+        return {"sampling_method": "ode", "cfg": 2.0, "sc_cfg": 1.0, "gamma": 0.0}
+    # 8/16/32: paper Appendix D.2; 64: release configuration.
+    # Other budgets are explicitly an extension of our benchmark.
+    return {"sampling_method": "sde", "cfg": 1.0, "sc_cfg": 3.0,
+            "gamma": 2.0 if steps <= 16 else 1.5 if steps == 32 else 1.0}
+
+
+def manifest(root):
+    path = Path(root) / "data/elf/assets.json"
+    if not path.is_file():
+        raise FileNotFoundError(
+            f"ELF asset manifest is missing: {path}. "
+            "First complete `python scripts/prepare_elf.py assets` successfully; "
+            "a failed download does not publish assets.json. Do not create it manually."
+        )
+    data = json.loads(path.read_text())
+    if data["lock_sha256"] != sha256(Path(root) / "artifacts/elf_lock.json"):
+        raise ValueError("ELF lock changed since asset preparation")
+    return data
+
+
+def asset(root, name):
+    record = manifest(root)["assets"][name]
+    directory = Path(root) / record["path"]
+    for relative, digest in record["sha256"].items():
+        if sha256(directory / relative) != digest:
+            raise ValueError(f"ELF asset changed: {name}/{relative}")
+    return directory
+
+
+def offline():
+    os.environ["HF_HUB_OFFLINE"] = "1"
+    os.environ["TRANSFORMERS_OFFLINE"] = "1"
+    os.environ["WANDB_MODE"] = "disabled"

scripts/evaluate_4gpu.sh

@@ -0,0 +1,48 @@
+#!/usr/bin/env bash
+# Four-GPU local quality evaluation over completed production samples.
+set -uo pipefail
+
+SCRIPT_DIR="$(CDPATH= cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd -P)"
+DLB_ROOT="${DLB_ROOT:-$(CDPATH= cd -- "$SCRIPT_DIR/.." && pwd -P)}"
+PYTHON_BIN="${DLB_PYTHON:-python}"
+MATRIX="${DLB_MATRIX:-}"
+CUSTOM_MATRIX=0
+if [[ -n "$MATRIX" ]]; then CUSTOM_MATRIX=1; fi
+METRICS="gen_ppl,entropy,self_bleu"
+GPUS="${DLB_GPUS:-0,1,2,3}"
+MAX_JOBS="${DLB_MAX_JOBS:-4}"
+DRY_RUN=()
+
+usage() {
+  printf 'Usage: %s [--root PATH] [--matrix PATH] [--metrics LIST] [--gpus LIST] [--max-jobs N] [--dry-run]\n' "$0"
+}
+
+while (($#)); do
+  case "$1" in
+    --root) DLB_ROOT="$2"; shift 2 ;;
+    --matrix) MATRIX="$2"; CUSTOM_MATRIX=1; shift 2 ;;
+    --metrics) METRICS="$2"; shift 2 ;;
+    --gpus) GPUS="$2"; shift 2 ;;
+    --max-jobs) MAX_JOBS="$2"; shift 2 ;;
+    --dry-run) DRY_RUN=(--dry-run); shift ;;
+    --help) usage; exit 0 ;;
+    *) echo "unknown argument: $1" >&2; usage; exit 2 ;;
+  esac
+done
+
+DLB_ROOT="$(CDPATH= cd -- "$DLB_ROOT" && pwd -P)"
+if [[ -z "$MATRIX" ]]; then MATRIX="$DLB_ROOT/results/matrix/generation.tsv"; fi
+export DLB_ROOT
+mkdir -p "$DLB_ROOT/results/matrix" "$DLB_ROOT/results/logs"
+if (( CUSTOM_MATRIX )); then
+  if [[ ! -f "$MATRIX" ]]; then
+    echo "custom matrix does not exist: $MATRIX" >&2
+    exit 2
+  fi
+else
+  PYTHONPATH="$DLB_ROOT/src${PYTHONPATH:+:$PYTHONPATH}" "$PYTHON_BIN" -m dlb.matrix \
+    --root "$DLB_ROOT" --output "$MATRIX" --sample-count 1024 --seed 42 || exit $?
+fi
+exec env PYTHONPATH="$DLB_ROOT/src${PYTHONPATH:+:$PYTHONPATH}" "$PYTHON_BIN" -m dlb.gpu_matrix \
+  --root "$DLB_ROOT" --stage evaluate --matrix "$MATRIX" \
+  --gpus "$GPUS" --max-jobs "$MAX_JOBS" --metrics "$METRICS" "${DRY_RUN[@]}"

scripts/evaluate_all.sh

@@ -0,0 +1,47 @@
+#!/usr/bin/env bash
+# Server-side quality evaluation over completed production samples.
+set -uo pipefail
+
+SCRIPT_DIR="$(CDPATH= cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd -P)"
+DLB_ROOT="${DLB_ROOT:-$(CDPATH= cd -- "$SCRIPT_DIR/.." && pwd -P)}"
+PYTHON_BIN="${DLB_EVAL_PYTHON:-${DLB_PYTHON:-python}}"
+MATRIX="${DLB_MATRIX:-}"
+METRICS="gen_ppl,entropy,self_bleu"
+
+usage() {
+  printf 'Usage: %s [--root PATH] [--matrix PATH] [--metrics LIST]\n' "$0"
+}
+while (($#)); do
+  case "$1" in
+    --root) DLB_ROOT="$2"; shift 2 ;;
+    --matrix) MATRIX="$2"; shift 2 ;;
+    --metrics) METRICS="$2"; shift 2 ;;
+    --help) usage; exit 0 ;;
+    *) echo "unknown argument: $1" >&2; usage; exit 2 ;;
+  esac
+done
+DLB_ROOT="$(CDPATH= cd -- "$DLB_ROOT" && pwd -P)"
+if [[ -z "$MATRIX" ]]; then MATRIX="$DLB_ROOT/results/matrix/generation.tsv"; fi
+export DLB_ROOT
+PYTHONPATH="$DLB_ROOT/src${PYTHONPATH:+:$PYTHONPATH}" "$PYTHON_BIN" -m dlb.matrix \
+  --root "$DLB_ROOT" --output "$MATRIX" --sample-count 1024 --seed 42 || exit $?
+
+mkdir -p "$DLB_ROOT/results/logs"
+failures="$DLB_ROOT/results/logs/evaluation_failures.tsv"
+printf 'task_id\tmodel\tdataset\tsteps\texit_code\n' >"$failures"
+status=0
+while IFS=$'\t' read -r task_id category model dataset steps sample_count seed environment adapter source provenance sample_dir metrics_path timing_path; do
+  [[ "$task_id" == "task_id" || -z "$task_id" ]] && continue
+  mkdir -p "$(dirname -- "$metrics_path")"
+  echo "EVALUATE $task_id"
+  PYTHONPATH="$DLB_ROOT/src${PYTHONPATH:+:$PYTHONPATH}" "$PYTHON_BIN" -m evaluation.evaluate \
+    --root "$DLB_ROOT" --samples "$sample_dir/samples.jsonl" \
+    --metrics "$METRICS" --dataset "$dataset" --output "$metrics_path"
+  exit_code=$?
+  if ((exit_code != 0)); then
+    printf '%s\t%s\t%s\t%s\t%s\n' "$task_id" "$model" "$dataset" "$steps" "$exit_code" >>"$failures"
+    status=1
+  fi
+done < <(tail -n +3 "$MATRIX")
+if [[ ! -s "$failures" ]]; then rm -f "$failures"; fi
+exit "$status"

scripts/evaluate_conditional_4gpu.sh

@@ -0,0 +1,48 @@
+#!/usr/bin/env bash
+# Four-GPU local quality evaluation over completed C64 conditional samples.
+set -uo pipefail
+
+SCRIPT_DIR="$(CDPATH= cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd -P)"
+DLB_ROOT="${DLB_ROOT:-$(CDPATH= cd -- "$SCRIPT_DIR/.." && pwd -P)}"
+PYTHON_BIN="${DLB_PYTHON:-python}"
+MATRIX="${DLB_CONDITIONAL_MATRIX:-}"
+CUSTOM_MATRIX=0
+if [[ -n "$MATRIX" ]]; then CUSTOM_MATRIX=1; fi
+METRICS="conditional_gen_ppl,mauve_suffix,entropy,self_bleu,prefix_exact_match"
+GPUS="${DLB_GPUS:-0,1,2,3}"
+MAX_JOBS="${DLB_MAX_JOBS:-4}"
+DRY_RUN=()
+
+usage() {
+  printf 'Usage: %s [--root PATH] [--matrix PATH] [--metrics LIST] [--gpus LIST] [--max-jobs N] [--dry-run]\n' "$0"
+}
+
+while (($#)); do
+  case "$1" in
+    --root) DLB_ROOT="$2"; shift 2 ;;
+    --matrix) MATRIX="$2"; CUSTOM_MATRIX=1; shift 2 ;;
+    --metrics) METRICS="$2"; shift 2 ;;
+    --gpus) GPUS="$2"; shift 2 ;;
+    --max-jobs) MAX_JOBS="$2"; shift 2 ;;
+    --dry-run) DRY_RUN=(--dry-run); shift ;;
+    --help) usage; exit 0 ;;
+    *) echo "unknown argument: $1" >&2; usage; exit 2 ;;
+  esac
+done
+
+DLB_ROOT="$(CDPATH= cd -- "$DLB_ROOT" && pwd -P)"
+if [[ -z "$MATRIX" ]]; then MATRIX="$DLB_ROOT/results/conditional/matrix/generation.tsv"; fi
+export DLB_ROOT
+mkdir -p "$DLB_ROOT/results/conditional/matrix" "$DLB_ROOT/results/logs"
+if (( CUSTOM_MATRIX )); then
+  if [[ ! -f "$MATRIX" ]]; then
+    echo "custom matrix does not exist: $MATRIX" >&2
+    exit 2
+  fi
+else
+  PYTHONPATH="$DLB_ROOT/src${PYTHONPATH:+:$PYTHONPATH}" "$PYTHON_BIN" -m dlb.conditional_matrix \
+    --root "$DLB_ROOT" --output "$MATRIX" || exit $?
+fi
+exec env PYTHONPATH="$DLB_ROOT/src${PYTHONPATH:+:$PYTHONPATH}" "$PYTHON_BIN" -m dlb.gpu_matrix \
+  --root "$DLB_ROOT" --stage evaluate --matrix "$MATRIX" \
+  --gpus "$GPUS" --max-jobs "$MAX_JOBS" --metrics "$METRICS" "${DRY_RUN[@]}"

scripts/evaluate_elf.py

@@ -0,0 +1,154 @@
+#!/usr/bin/env python3
+"""Score independent ELF artifacts with pinned, offline evaluation assets."""
+from __future__ import annotations
+
+import argparse
+from collections import defaultdict
+from dataclasses import asdict
+import json
+import math
+from pathlib import Path
+import statistics
+import sys
+
+from elf_common import ROOT, asset, manifest, offline, read_jsonl, require_server, sha256, write_json
+
+
+def validate_records(rows, meta):
+    if len(rows) != meta["sample_count"]:
+        raise ValueError("Sample count differs from generation manifest")
+    n = meta["prompt_count"]
+    expected = [(i, 0) for i in range(n)]
+    if meta.get("diversity"):
+        expected += [(i, c) for c in range(1, 5) for i in range(min(256, n))]
+    actual = [(r["prompt_id"], r["completion_id"]) for r in rows]
+    if actual != expected or [r["id"] for r in rows] != list(range(len(rows))):
+        raise ValueError("Missing/duplicated/out-of-order prompt completions")
+
+
+def main():
+    p = argparse.ArgumentParser(description=__doc__)
+    p.add_argument("--run", type=Path, required=True)
+    p.add_argument("--root", type=Path, default=ROOT)
+    p.add_argument("--batch-size", type=int, default=8)
+    p.add_argument("--mauve", action="store_true")
+    args = p.parse_args()
+    require_server()
+    offline()
+    root = args.root.resolve()
+    sys.path[:0] = [str(root), str(root / "src")]
+    meta = json.loads((args.run / "generation.json").read_text())
+    samples = args.run / "samples.jsonl"
+    if sha256(samples) != meta["samples_sha256"]:
+        raise ValueError("Sample file changed since generation")
+    if sha256(root / "data/elf/assets.json") != meta["assets_manifest_sha256"]:
+        raise ValueError("Asset manifest changed since generation")
+    rows = read_jsonl(samples)
+    validate_records(rows, meta)
+    primary = [r for r in rows if r["completion_id"] == 0]
+    result = {"schema": "dlb-elf-metrics-v1", "task": meta["task"],
+              "generation_manifest_sha256": sha256(args.run / "generation.json"),
+              "samples_sha256": sha256(samples), "sample_count": len(primary),
+              "evaluator_sha256": sha256(Path(__file__)), "metrics": {}}
+    metrics = result["metrics"]
+    if meta["task"] in ("wmt14", "xsum"):
+        import sacrebleu
+        from rouge_score import rouge_scorer
+        hypotheses, refs = [r["generated"] for r in primary], [r["reference"] for r in primary]
+        # Same metric options as official utils/metrics_utils.py.
+        bleu = sacrebleu.metrics.BLEU(lowercase=True, effective_order=True)
+        metrics["bleu"] = bleu.corpus_score(hypotheses, [refs]).score
+        metrics["bleu_signature"] = str(bleu.get_signature())
+        metrics["empty_fraction"] = sum(not x.strip() for x in hypotheses) / len(hypotheses)
+        scorer = rouge_scorer.RougeScorer(["rouge1", "rouge2", "rougeL"], use_stemmer=True)
+        scores = [scorer.score(ref, hyp) for ref, hyp in zip(refs, hypotheses)]
+        for key in ("rouge1", "rouge2", "rougeL"):
+            values = [s[key].fmeasure * 100 for s in scores]
+            metrics[key] = statistics.fmean(values)
+            metrics[key + "_sem"] = statistics.pstdev(values) / math.sqrt(len(values))
+        metrics["reference_policy"] = "full raw reference; no target truncation; ROUGE F1 stemmer=true"
+    else:
+        import torch
+        from transformers import AutoTokenizer, AutoModelForCausalLM
+        from evaluation.generative_perplexity import compute_gen_ppl, aggregate_nll
+        from evaluation.conditional_perplexity import PromptExcludedCausalLMScorer
+        from evaluation.unigram_entropy import unigram_entropy
+        from evaluation.self_bleu import compute_self_bleu
+        gpt_path = asset(root, "gpt2")
+        scorer_path = asset(root, "gpt2-large")
+        tok = AutoTokenizer.from_pretrained(str(gpt_path), local_files_only=True)
+        tok.pad_token = tok.eos_token
+        tok.padding_side = "right"
+        tok.truncation_side = "right"
+        tokenized = [tok(r["generated"], add_special_tokens=False)["input_ids"] for r in rows]
+        lengths = [len(ids) for ids in tokenized[:len(primary)]]
+        metrics["gpt2_length_mean"] = statistics.fmean(lengths)
+        metrics["gpt2_length_min"] = min(lengths)
+        metrics["empty_fraction"] = sum(n == 0 for n in lengths) / len(lengths)
+        metrics["below_64_tokens_fraction"] = sum(n < 64 for n in lengths) / len(lengths)
+        metrics["above_1024_tokens_fraction"] = sum(n > 1024 for n in lengths) / len(lengths)
+        limit = 64 if meta["task"] == "owt-prefix" else 1024
+        tokenized = [ids[:limit] for ids in tokenized]
+        # Never improve apparent PPL by silently removing degenerate generations.
+        if any(len(ids) < 2 for ids in tokenized):
+            result["valid"] = False
+            result["reason"] = "At least one completion has fewer than two GPT-2 tokens; no samples were dropped"
+            write_json(args.run / "metrics.json", result)
+            raise ValueError(result["reason"])
+        metrics["entropy_gpt2_nats"] = statistics.fmean(unigram_entropy(ids) for ids in tokenized[:len(primary)])
+        result["evaluator_assets"] = {k: manifest(root)["assets"][k] for k in ("gpt2", "gpt2-large")}
+        model = AutoModelForCausalLM.from_pretrained(str(scorer_path), local_files_only=True).cuda().eval()
+        if meta["task"] == "owt":
+            ppl = compute_gen_ppl([r["generated"] for r in primary], model, tok,
+                                  batch_size=args.batch_size, device="cuda",
+                                  model_revision=result["evaluator_assets"]["gpt2-large"]["revision"],
+                                  tokenizer_revision=result["evaluator_assets"]["gpt2"]["revision"])
+            metrics["generative_ppl"] = asdict(ppl)
+        else:
+            scorer = PromptExcludedCausalLMScorer(model, device="cuda")
+            def conditional_ppl(targets):
+                parts = []
+                for start in range(0, len(primary), args.batch_size):
+                    batch = primary[start:start + args.batch_size]
+                    combined, loss_masks = [], []
+                    for j, row in enumerate(batch):
+                        prefix = tok(row["input"], add_special_tokens=False)["input_ids"]
+                        target = targets[start + j]
+                        if not prefix or len(prefix) + len(target) > 1024:
+                            raise ValueError("Conditional PPL input exceeds context or has empty prefix")
+                        combined.append(prefix + target)
+                        loss_masks.append([0] * len(prefix) + [1] * len(target))
+                    width = max(map(len, combined))
+                    attn = [[1] * len(x) + [0] * (width - len(x)) for x in combined]
+                    masks = [x + [0] * (width - len(x)) for x in loss_masks]
+                    ids = [x + [tok.pad_token_id] * (width - len(x)) for x in combined]
+                    parts.append(scorer.score_prompt_excluded_batch(ids, attn, masks))
+                return asdict(aggregate_nll(parts))
+            metrics["conditional_ppl"] = conditional_ppl(tokenized[:len(primary)])
+            references = [tok(r["reference"], add_special_tokens=False)["input_ids"][:64] for r in primary]
+            metrics["reference_conditional_ppl"] = conditional_ppl(references)
+            metrics["conditional_ppl_policy"] = "separate GPT-2 prefix/suffix tokenization; prefix targets masked"
+            metrics["entropy_all_completions_gpt2_nats"] = statistics.fmean(unigram_entropy(ids) for ids in tokenized)
+            if meta.get("diversity"):
+                groups = defaultdict(list)
+                for row, ids in zip(rows, tokenized):
+                    if row["prompt_id"] < min(256, meta["prompt_count"]):
+                        groups[row["prompt_id"]].append(" ".join(map(str, ids)))
+                metrics["grouped_self_bleu"] = statistics.fmean(compute_self_bleu(group).score for group in groups.values())
+            if args.mauve:
+                import mauve
+                del scorer, model
+                torch.cuda.empty_cache()
+                score = mauve.compute_mauve(
+                    p_text=[tok.decode(ids) for ids in tokenized[:len(primary)]],
+                    q_text=[tok.decode(ids) for ids in references], featurize_model_name=str(scorer_path),
+                    device_id=0, max_text_length=64, seed=42, verbose=False)
+                metrics["mauve"] = float(score.mauve)
+        metrics["entropy_policy"] = "mean per-sample natural-log entropy on retokenized GPT-2 suffix/text; same PPL length cap"
+    result["valid"] = True
+    write_json(args.run / "metrics.json", result)
+    print(json.dumps(metrics, indent=2))
+
+
+if __name__ == "__main__":
+    main()

scripts/evaluate_elf_owt_baseline.py

@@ -0,0 +1,175 @@
+#!/usr/bin/env python3
+"""Rescore public ELF OWT outputs against the existing C64 baseline (no generation)."""
+from __future__ import annotations
+
+import argparse
+import csv
+from dataclasses import dataclass, asdict
+import json
+from pathlib import Path
+import statistics
+import sys
+
+from elf_common import ROOT, asset, manifest, offline, read_jsonl, require_server, sha256, write_json
+from run_elf_remaining import build_matrix, reusable_result
+
+
+@dataclass
+class C64Record:
+    """Token-record interface consumed by the existing baseline evaluator."""
+    prompt_id: int
+    completion_id: int
+    prefix_token_ids: list[int]
+    continuation_token_ids: list[int]
+    reference_token_ids: list[int]
+
+
+def c64_records(rows, prompts, tokenized, tokenizer, *, allow_short=False):
+    by_id = {p['prompt_id']: p for p in prompts}
+    records = []
+    for row, ids in zip(rows, tokenized, strict=True):
+        source = by_id[row['prompt_id']]
+        prefix, reference = source['prefix_token_ids'], source['reference_token_ids']
+        if len(prefix) != 64 or len(reference) != 64:
+            raise ValueError('Canonical prompt/reference must contain exactly 64 GPT-2 tokens')
+        # Prove the saved ELF condition/reference came from this exact prompt.
+        if row['input'] != tokenizer.decode(prefix, skip_special_tokens=False):
+            raise ValueError('Saved ELF input differs from canonical C64 prompt')
+        if row['reference'] != tokenizer.decode(reference, skip_special_tokens=False):
+            raise ValueError('Saved ELF reference differs from canonical C64 reference')
+        if not ids or (len(ids) < 64 and not allow_short):
+            raise ValueError('Short response cannot satisfy the 64-token scoring contract')
+        records.append(C64Record(row['prompt_id'], row['completion_id'], prefix, ids[:64], reference))
+    return records
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument('--root', type=Path, default=ROOT)
+    parser.add_argument('--task', choices=['owt', 'owt-prefix'], required=True)
+    parser.add_argument('--output', type=Path)
+    parser.add_argument('--batch-size', type=int, default=8)
+    parser.add_argument('--steps', nargs='+', type=int, choices=[1, 2, 4, 8, 16, 32, 1024])
+    parser.add_argument('--short-response-policy', choices=['strict', 'observed'], default='strict',
+                        help='observed: separately report all nonempty responses up to 64 GPT-2 tokens')
+    parser.add_argument('--plan', action='store_true')
+    args = parser.parse_args()
+    if args.batch_size <= 0:
+        parser.error('--batch-size must be positive')
+    observed = args.short_response_policy == 'observed'
+    if observed and args.task != 'owt-prefix':
+        parser.error('observed short-response policy is only for owt-prefix')
+    if args.steps and len(args.steps) != len(set(args.steps)):
+        parser.error('duplicate steps')
+    root = args.root.resolve()
+    sys.path[:0] = [str(root), str(root / 'src')]
+    output = args.output or root / 'results' / (f'elf-owt-baseline-eval-{args.task}' + ('-observed' if observed else ''))
+    assets = manifest(root)['assets']
+    asset_hash = sha256(root / 'data/elf/assets.json')
+    cells, blocked = build_matrix(root, [args.task])
+    if blocked:
+        raise ValueError(f'Invalid inputs: {blocked}')
+    if args.steps:
+        cells = [cell for cell in cells if cell.steps in args.steps]
+    runs = []
+    for cell in cells:
+        found, rejected = reusable_result(root, cell, asset_hash, assets)
+        if not found:
+            raise ValueError(f'Missing verified generation: {cell.key}; first run run_elf_remaining.py; rejected={rejected}')
+        path = Path(found['run'])
+        runs.append((cell, path))
+        print(f'RESCORE {cell.key}: {path}', flush=True)
+    if args.plan:
+        print(f'{len(runs)} cells; no generation, no training; output={output}')
+        return
+    require_server()
+    offline()
+    from transformers import AutoTokenizer, AutoModelForCausalLM
+    from evaluation.generative_perplexity import compute_gen_ppl
+    from evaluation.conditional_perplexity import compute_conditional_gen_ppl, conditional_texts
+    from evaluation.conditional_evaluate import _gpt2_tokenized_self_bleu
+    from evaluation.unigram_entropy import unigram_entropy
+    tok = AutoTokenizer.from_pretrained(str(asset(root, 'gpt2')), local_files_only=True)
+    tok.pad_token = tok.eos_token
+    tok.padding_side = tok.truncation_side = 'right'
+    prompts = read_jsonl(root / 'data/conditional/owt-c64/prompts.jsonl') if args.task == 'owt-prefix' else []
+    dependencies = [Path(__file__), root / 'evaluation/generative_perplexity.py',
+                    root / 'evaluation/conditional_perplexity.py', root / 'evaluation/conditional_evaluate.py',
+                    root / 'evaluation/unigram_entropy.py', root / 'evaluation/self_bleu.py']
+    evaluator = {p.name: sha256(p) for p in dependencies}
+    model = None
+    summary = []
+    for cell, path in runs:
+        destination = output / f'steps_{cell.steps}' / 'metrics.json'
+        provenance = {'protocol': 'elf_owt_public_baseline_observed_v1' if observed else 'elf_owt_public_baseline_eval_v1', 'task': args.task,
+                      'steps': cell.steps, 'source_run': str(path),
+                      'samples_sha256': sha256(path / 'samples.jsonl'),
+                      'generation_manifest_sha256': sha256(path / 'generation.json'),
+                      'assets_manifest_sha256': asset_hash, 'evaluator': evaluator,
+                      'batch_size': args.batch_size,
+                      'prompts_sha256': sha256(root / 'data/conditional/owt-c64/prompts.jsonl') if prompts else None}
+        cached = json.loads(destination.read_text()) if destination.exists() else {}
+        if cached and cached.get('protocol') != provenance['protocol']:
+            raise ValueError(f'Different protocol already exists at {destination}; use a separate --output')
+        if all(cached.get(k) == v for k, v in provenance.items()) and type(cached.get('valid')) is bool:
+            result = cached
+            print(f'SKIP verified evaluation: {destination}', flush=True)
+        else:
+            rows = read_jsonl(path / 'samples.jsonl')
+            ids = [tok(row['generated'], add_special_tokens=False)['input_ids'] for row in rows]
+            lengths = list(map(len, ids))
+            limit = 64 if prompts else 1024
+            short = [row['id'] for row, tokens in zip(rows, ids) if len(tokens) < (64 if prompts else 2)]
+            invalid = [row['id'] for row, tokens in zip(rows, ids) if not tokens] if observed else short
+            result = {**provenance, 'sample_count': len(rows), 'prompt_count': 1024,
+                      'metrics': {}, 'valid': not invalid,
+                      'short_response_policy': args.short_response_policy,
+                      'invalid_sample_ids': invalid,
+                      'length_diagnostics': {'mean': statistics.fmean(lengths), 'min': min(lengths),
+                                             'max': max(lengths), 'below_64_count': sum(n < 64 for n in lengths),
+                                             'above_1024_count': sum(n > 1024 for n in lengths)},
+                      'short_sample_ids': short,
+                      'native_tokenizer': 't5-small', 'scoring_tokenizer': 'gpt2',
+                      'scoring_limit': limit, 'completion_scope': 'all_saved_completions',
+                      'output_policy': 'released ELF EOS-trimmed decoded text; GPT-2 retokenized; no fabricated padding'}
+            if invalid:
+                result['reason'] = ('Empty outputs' if observed else 'Short outputs') + ': cannot score the requested contract; none dropped, padded, or resampled'
+            else:
+                sliced = [tokens[:limit] for tokens in ids]
+                result['metrics']['entropy_gpt2_nats'] = statistics.fmean(unigram_entropy(tokens) for tokens in sliced)
+                records = c64_records(rows, prompts, ids, tok, allow_short=observed) if prompts else None
+                if model is None:
+                    model = AutoModelForCausalLM.from_pretrained(str(asset(root, 'gpt2-large')), local_files_only=True).cuda().eval()
+                kwargs = dict(batch_size=args.batch_size, device='cuda',
+                              model_revision=assets['gpt2-large']['revision'], tokenizer_revision=assets['gpt2']['revision'])
+                if prompts:
+                    ppl = compute_conditional_gen_ppl(records, model, tok, tok, allow_short_continuations=observed, **kwargs)
+                    result['metrics']['conditional_ppl'] = asdict(ppl)
+                    texts = conditional_texts(records, tok, allow_short_continuations=observed)
+                    result['metrics']['grouped_self_bleu'] = _gpt2_tokenized_self_bleu(records, [t.generated_suffix for t in texts], tok)
+                    result['ppl_policy'] = 'existing compute_conditional_gen_ppl; canonical prefix IDs; decode joined IDs then retokenize; prompt loss excluded'
+                else:
+                    ppl = compute_gen_ppl([row['generated'] for row in rows], model, tok, **kwargs)
+                    result['metrics']['generative_ppl'] = asdict(ppl)
+                if ppl.sample_count != len(rows):
+                    raise ValueError('PPL silently changed completion count')
+            write_json(destination, result)
+        metrics = result['metrics']
+        ppl = metrics.get('conditional_ppl', metrics.get('generative_ppl', {})).get('perplexity')
+        summary.append({'task': args.task, 'steps': cell.steps, 'protocol': result['protocol'], 'sample_count': result['sample_count'],
+                        'valid': result['valid'], 'ppl': ppl, 'entropy': metrics.get('entropy_gpt2_nats'),
+                        'short_count': len(result['short_sample_ids']), 'metrics_path': str(destination)})
+        print(json.dumps(summary[-1]), flush=True)
+    output.mkdir(parents=True, exist_ok=True)
+    summary_name = 'summary' + ('-steps-' + '-'.join(str(c.steps) for c, _ in runs) if args.steps else '') + '.csv'
+    with (output / summary_name).open('w') as handle:
+        writer = csv.DictWriter(handle, fieldnames=list(summary[0]))
+        writer.writeheader()
+        writer.writerows(summary)
+    print(f'SUMMARY: {output / summary_name}', flush=True)
+    if any(not row['valid'] for row in summary):
+        raise SystemExit(2)
+
+
+if __name__ == '__main__':
+    main()

scripts/fetch_checkpoints.py

@@ -0,0 +1,67 @@
+#!/usr/bin/env python3
+"""Enumerate or fetch pinned public checkpoint resources."""
+
+from __future__ import annotations
+
+import argparse
+from pathlib import Path
+import platform
+import sys
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--root", type=Path, default=Path(__file__).resolve().parents[1])
+    parser.add_argument("--config", type=Path)
+    parser.add_argument("--registry", type=Path)
+    parser.add_argument("--dry-run", action="store_true")
+    parser.add_argument("--all-public", action="store_true")
+    return parser.parse_args()
+
+
+def main() -> int:
+    args = parse_args()
+    root = args.root.resolve()
+    project_root = Path(__file__).resolve().parents[1]
+    sys.path.insert(0, str(project_root / "src"))
+    from dlb.checkpoints import (
+        fetch_all_resources,
+        load_checkpoint_manifest,
+        require_server_platform,
+        validate_checkpoint_coverage,
+    )
+    from dlb.registry import load_registry
+
+    manifest_path = (args.config or root / "artifacts" / "checkpoints.yaml").resolve()
+    registry_path = (args.registry or root / "configs" / "experiments.yaml").resolve()
+    manifest = load_checkpoint_manifest(manifest_path)
+    registry = load_registry(registry_path)
+    validate_checkpoint_coverage(registry, manifest)
+    if args.dry_run:
+        for resource_id, resource in manifest.resources.items():
+            print(
+                f"RESOURCE {resource_id} {resource.backend} {resource.provenance} "
+                f"checkpoints/{resource.destination} teacher={resource.teacher_family}"
+            )
+        for recipe_id, recipe in manifest.recipes.items():
+            print(
+                f"RECIPE {recipe_id} {recipe.model}/{recipe.dataset} {recipe.provenance} "
+                f"{recipe.output} teacher={recipe.teacher_family}"
+            )
+        return 0
+    if not args.all_public:
+        print("ERROR: pass --all-public to acknowledge all public checkpoint downloads", file=sys.stderr)
+        return 2
+    try:
+        require_server_platform(platform.system())
+    except RuntimeError as error:
+        print(f"ERROR: {error}", file=sys.stderr)
+        return 2
+    lock = fetch_all_resources(root, manifest_path, manifest)
+    for resource_id, record in lock["resources"].items():
+        print(f"{str(record['status']).upper()} {resource_id} {record['destination']}")
+    return 1 if any(record["status"] != "downloaded" for record in lock["resources"].values()) else 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

scripts/fetch_data.py

@@ -0,0 +1,201 @@
+#!/usr/bin/env python3
+"""Download pinned datasets and model/tokenizer snapshots with cache resume."""
+
+from __future__ import annotations
+
+import argparse
+from datetime import datetime, timezone
+import os
+from pathlib import Path
+import shutil
+import sys
+
+import yaml
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--root", type=Path, default=Path(__file__).resolve().parents[1])
+    parser.add_argument("--config", type=Path)
+    parser.add_argument("--dry-run", action="store_true")
+    parser.add_argument("--allow-low-disk", action="store_true")
+    return parser.parse_args()
+
+
+def load_configuration(root: Path, config_path: Path | None) -> dict[str, object]:
+    path = config_path or root / "artifacts" / "data.yaml"
+    return yaml.safe_load(path.read_text(encoding="utf-8"))
+
+
+def target_lines(configuration: dict[str, object]) -> list[str]:
+    datasets = configuration["datasets"]
+    models = configuration["models"]
+    return [
+        *[
+            f"DATASET {name} {datasets[name]['repo_id']}@{datasets[name]['revision']}"
+            for name in ("lm1b", "owt")
+        ],
+        *[
+            f"MODEL {name}@{models[name]}"
+            for name in ("bert-base-uncased", "gpt2", "gpt2-large")
+        ],
+    ]
+
+
+def directory_size(path: Path) -> int:
+    return sum(candidate.stat().st_size for candidate in path.rglob("*") if candidate.is_file())
+
+
+def _relative_cache_path(path: Path, root: Path) -> str:
+    try:
+        return path.resolve().relative_to(root).as_posix()
+    except ValueError:
+        return path.resolve().as_posix()
+
+
+def main() -> int:
+    args = parse_args()
+    root = args.root.resolve()
+    configuration = load_configuration(root, args.config)
+    for line in target_lines(configuration):
+        print(line)
+    if args.dry_run:
+        return 0
+
+    hf_home = root / "data" / "raw" / "huggingface"
+    datasets_cache = hf_home / "datasets"
+    hub_cache = hf_home / "hub"
+    hf_home.mkdir(parents=True, exist_ok=True)
+    os.environ["HF_HOME"] = str(hf_home)
+    os.environ["HF_DATASETS_CACHE"] = str(datasets_cache)
+
+    from datasets import load_dataset
+    from huggingface_hub import snapshot_download
+
+    sys.path.insert(0, str(root / "src"))
+    from dlb.data import disk_preflight_evidence, validate_download_manifest
+    from dlb.io import atomic_json_write
+
+    dataset_records: dict[str, object] = {}
+    owt_disk_preflight = None
+    for name in ("lm1b", "owt"):
+        specification = configuration["datasets"][name]
+        if name == "owt":
+            free_bytes = shutil.disk_usage(root).free
+            owt_disk_preflight = disk_preflight_evidence(
+                free_bytes, allow_low_disk=args.allow_low_disk
+            )
+            if owt_disk_preflight["override_used"]:
+                print(
+                    f"WARNING low disk override: {free_bytes} bytes free, required 55 GiB",
+                    file=sys.stderr,
+                )
+        if name == "lm1b":
+            source_metadata_snapshot = snapshot_download(
+                repo_id=specification["repo_id"],
+                repo_type="dataset",
+                revision=specification["revision"],
+                cache_dir=hub_cache,
+            )
+            dataset_snapshot = snapshot_download(
+                repo_id=specification["repo_id"],
+                repo_type="dataset",
+                revision=specification["parquet_revision"],
+                cache_dir=hub_cache,
+                allow_patterns="plain_text/**/*.parquet",
+            )
+            snapshot_path = Path(dataset_snapshot)
+            loaded = load_dataset(
+                "parquet",
+                data_files={
+                    "train": [
+                        str(path)
+                        for path in sorted(
+                            (snapshot_path / "plain_text" / "train").glob("*.parquet")
+                        )
+                    ],
+                    "test": [
+                        str(path)
+                        for path in sorted(
+                            (snapshot_path / "plain_text" / "test").glob("*.parquet")
+                        )
+                    ],
+                },
+                cache_dir=str(datasets_cache),
+                download_mode="reuse_dataset_if_exists",
+            )
+            archive_specification = configuration["lm1b_archive"]
+            if len(loaded["train"]) != int(archive_specification["train_documents"]):
+                raise RuntimeError("LM1B training document count differs from pinned metadata")
+            if len(loaded["test"]) != int(archive_specification["test_documents"]):
+                raise RuntimeError("LM1B test document count differs from pinned metadata")
+        else:
+            source_metadata_snapshot = None
+            dataset_snapshot = None
+            loaded = load_dataset(
+                specification["repo_id"],
+                specification["config"],
+                revision=specification["revision"],
+                cache_dir=str(datasets_cache),
+                download_mode="reuse_dataset_if_exists",
+            )
+            if len(loaded["train"]) != int(specification["documents"]):
+                raise RuntimeError("OpenWebText document count differs from pinned metadata")
+        dataset_records[name] = {
+            "repo_id": specification["repo_id"],
+            "source_revision": specification["revision"],
+            "cache_files": {
+                split: [
+                    _relative_cache_path(Path(entry["filename"]), root)
+                    for entry in value.cache_files
+                ]
+                for split, value in loaded.items()
+            },
+            "split_rows": {split: len(value) for split, value in loaded.items()},
+            "dataset_snapshot": (
+                _relative_cache_path(Path(dataset_snapshot), root) if dataset_snapshot else None
+            ),
+            "source_metadata_snapshot": (
+                _relative_cache_path(Path(source_metadata_snapshot), root)
+                if source_metadata_snapshot
+                else None
+            ),
+            "materialization_revision": specification.get("parquet_revision"),
+        }
+        print(f"DOWNLOADED DATASET {name} {dataset_records[name]['split_rows']}")
+
+    model_records: dict[str, object] = {}
+    for name in ("bert-base-uncased", "gpt2", "gpt2-large"):
+        revision = configuration["models"][name]
+        snapshot_path = Path(
+            snapshot_download(
+                repo_id=name,
+                revision=revision,
+                cache_dir=hub_cache,
+            )
+        )
+        model_records[name] = {
+            "repo_id": name,
+            "revision": revision,
+            "snapshot_path": snapshot_path.relative_to(root).as_posix(),
+            "size_bytes": directory_size(snapshot_path),
+        }
+        print(f"DOWNLOADED MODEL {name} {snapshot_path}")
+
+    manifest = {
+        "schema_version": 1,
+        "created_at": datetime.now(timezone.utc).isoformat().replace("+00:00", "Z"),
+        "hf_home": hf_home.relative_to(root).as_posix(),
+        "owt_disk_preflight": owt_disk_preflight,
+        "datasets": dataset_records,
+        "models": model_records,
+    }
+    for dataset_name in ("lm1b", "owt"):
+        validate_download_manifest(configuration, manifest, dataset_name)
+    atomic_json_write(root / "data" / "manifests" / "downloads.json", manifest)
+    print(f"RAW CACHE BYTES {directory_size(hf_home)}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

scripts/inspect_elf_progress.py

@@ -0,0 +1,84 @@
+#!/usr/bin/env python3
+"""Read-only ELF quality inventory; verify saved artifacts before reporting completion."""
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+
+from elf_common import ROOT, read_jsonl, sha256, write_json
+from evaluate_elf import validate_records
+
+OWT_MAIN_STEPS = (1, 2, 4, 8, 16, 32, 1024)
+
+
+def inspect_run(path):
+    row = {"run": str(path), "status": "incomplete_generation"}
+    try:
+        request = json.loads((path / "request.json").read_text())
+        row.update(task=request["task"], steps=request["steps"], seed=request["seed"],
+                   compile=request["compile"],
+                   split=(request.get("input") or {}).get(
+                       "split", "c64-text-t5" if request["task"] == "owt-prefix" else "unconditional"),
+                   protocol=request.get("protocol"),
+                   condition_policy=request.get("condition_tokenization_policy"))
+        gp = path / "generation.json"
+        if not gp.exists():
+            return row
+        gen = json.loads(gp.read_text())
+        for key in ("task", "steps", "seed", "compile"):
+            if gen[key] != request[key]:
+                raise ValueError(f"request/generation mismatch: {key}")
+        sample_hash = sha256(path / "samples.jsonl")
+        if gen["samples_sha256"] != sample_hash:
+            raise ValueError("samples hash mismatch")
+        validate_records(read_jsonl(path / "samples.jsonl"), gen)
+        row.update(sample_count=gen["sample_count"], prompt_count=gen["prompt_count"],
+                   status="needs_evaluation")
+        mp = path / "metrics.json"
+        if mp.exists():
+            metric = json.loads(mp.read_text())
+            if (metric["generation_manifest_sha256"] != sha256(gp)
+                    or metric["samples_sha256"] != sample_hash
+                    or metric["task"] != gen["task"]
+                    or metric["sample_count"] != gen["prompt_count"]):
+                raise ValueError("metrics provenance/count mismatch")
+            if type(metric.get("valid")) is not bool:
+                raise ValueError("missing boolean metric validity")
+            row.update(status="quality_valid" if metric["valid"] else "quality_invalid",
+                       metrics=metric["metrics"], reason=metric.get("reason"))
+        row["owt_main_grid"] = (
+            row["task"] == "owt" and row["steps"] in OWT_MAIN_STEPS
+            and row["prompt_count"] == 1024 and not gen.get("diversity")
+            and row["seed"] == 42 and not row["compile"]
+        )
+    except (OSError, ValueError, KeyError, TypeError) as error:
+        row.update(status="invalid_artifact", reason=str(error))
+    return row
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--root", type=Path, default=ROOT)
+    parser.add_argument("--output", type=Path)
+    args = parser.parse_args()
+    paths = sorted((args.root / "results").glob("elf*/quality/*/*/steps_*/seed_*/compile_*"))
+    rows = [inspect_run(path) for path in paths if path.is_dir()]
+    for row in rows:
+        metrics = row.get("metrics", {})
+        scores = {key: metrics[key] for key in ("bleu", "rouge1", "rouge2", "rougeL", "entropy_gpt2_nats") if key in metrics}
+        for key in ("generative_ppl", "conditional_ppl"):
+            if key in metrics:
+                scores[key] = metrics[key].get("perplexity")
+        print(json.dumps({k: v for k, v in row.items() if k != "metrics"} | {"scores": scores}, ensure_ascii=False))
+    present = {row["steps"] for row in rows if row.get("owt_main_grid") and row["status"] == "quality_valid"}
+    missing = [step for step in OWT_MAIN_STEPS if step not in present]
+    print(f"OWT primary grid {list(OWT_MAIN_STEPS)}; steps without a verified valid result: {missing}")
+    print("Old/no-EOS conditional runs and auxiliary steps remain listed; do not merge protocols. No GPU work performed.")
+    if args.output:
+        write_json(args.output, {"schema": "elf-progress-v1", "runs": rows,
+                                "owt_steps_without_valid_result": missing})
+
+
+if __name__ == "__main__":
+    main()