Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): ('imagefolder', {}), NamedSplit('test'): ('text', {})}
Error code:   FileFormatMismatchBetweenSplitsError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

craftax ReMDM planner — RL fine-tuning ablations (temporary)

Merged results of the 25-arm RL fine-tuning ablation suite, 3 seeds per arm.

Date 2026-08-21
Repo craftax-ReMDM-planner @ 9a45363
Config experiments/rl_finetuning/configs/ablations_final_craftax_classic_ucl.yaml
Env Craftax-Classic-Symbolic-v1
Checkpoint W&B myopic-planner/craftax-ReMDM-planner/8qw13bmd -> artifact Craftax-Classic-Symbolic-v1-policy-best:v2 (Orbax step 40370176)
Hardware 3x RTX 3090 Ti (UCL: shoveler-l, smew-l, wigeon-l)
Seeds 3 per arm

Produced by run_ablations.py --merge over three per-GPU shards; all three shards ran one config, so --merge pooled them without refusal.

Contents: results.json, diagnosis.md, 113 figures, 19 tables.

Temporary staging copy — not a release artefact.

pretrained_checkpoint/

The exact pretrained planner these 25 ablations fine-tune from.

W&B run myopic-planner/craftax-ReMDM-planner/8qw13bmd
Artifact Craftax-Classic-Symbolic-v1-policy-best:v2 (best-validation, not final)
Format Orbax checkpoint, step 40370176
Size 34 MB, 13 files

Restores with src/planners/model.restore_latest. The sibling artifact Craftax-Classic-Symbolic-v1-policy:v3 is the final-step checkpoint of the same run and is not what these ablations used.

inference/

--mode inference evaluation of the bundled checkpoint, run 2026-08-21 on a UCL RTX 3090 Ti (shoveler-l).

Command --mode inference --config configs/final_craftax_classic_ucl.yaml --override use_wandb=false
Protocol 32 agents x 10,000 steps, seed 42, diffusion_steps_eval 10
mean_score 3.256 (best 8.100)
mean episode length 161.2

Top achievement rates: collect_wood 0.97, collect_sapling 0.94, place_table 0.81, place_plant 0.50. Zero: eat_plant, defeat_skeleton, make_stone_sword.

Not comparable to results.json's pretrained_score

results.json records pretrained_score ~11.72 for this same checkpoint. The two numbers come from different samplers, not different settings:

--mode inference ablation suite eval
sampler sample_plan_inpainting sample_plan
replan cadence every env step every EVAL_REPLAN = 8 steps
actions per plan 1 (history-conditioned) 8
denoising steps DIFFUSION_STEPS_EVAL = 10 VAL_DIFFUSION_STEPS = 50

The denoising budget is not the cause: re-running inference with --override diffusion_steps_eval=50 gives 3.4 against 3.3 at 10 steps, a ~3% move against a 3.6x gap. The sampler and replan cadence account for the rest.

offline_bc_checkpoint/ + inference/craftax_offline_bc_final.json

An offline BC planner trained 2026-08-22, distinct from the DAgger checkpoint above. Added so this repo carries both arms of the comparison.

W&B run myopic-planner/craftax-ReMDM-planner/4oeqk0yl (…-Offline-Diffusion-BC-99M)
Training --mode offline --config configs/final_craftax_classic_ucl.yaml, seed 42
Expert Craftax-Classic-Symbolic-v1-PPO_RNN-1000M (rolled out live; --mode offline consumes no dataset)
Budget 99,942,400 env frames = 1525 updates x 512 envs x 128 steps; 97,600 gradient steps
Wall clock 56,283 s (~15.6 h) on a UCL RTX 3090 Ti (shoveler-l), SPS 1778
Format Orbax, step 99942400 (98 MB)
Also on W&B as Craftax-Classic-Symbolic-v1-policy:v4

Pass the directory, not the step subdirectory — CheckpointManager resolves the step itself.

Inference vs the DAgger checkpoint

Identical protocol both sides: 32 agents x 10,000 steps, seed 42, diffusion_steps_eval 10.

metric offline BC DAgger policy-best:v2
mean_score 3.8812 3.2562
best_score 7.1000 8.1000
mean_episode_length 139.59 161.16

Offline BC scores +0.625 mean over DAgger, though its single best episode is lower (7.10 vs 8.10) — it is more consistent rather than higher-ceilinged.

Achievement rates where the two differ:

achievement offline BC DAgger delta
make_wood_pickaxe 0.406 0.188 +0.219
make_wood_sword 0.156 0.031 +0.125
defeat_zombie 0.125 0.031 +0.094
make_stone_sword 0.094 0.000 +0.094
collect_sapling 1.000 0.938 +0.062
place_table 0.875 0.812 +0.062
place_plant 0.562 0.500 +0.062
collect_stone 0.219 0.156 +0.062
collect_drink 0.125 0.062 +0.062
make_stone_pickaxe 0.000 0.094 -0.094
eat_cow 0.062 0.125 -0.062
place_furnace 0.062 0.094 -0.031
place_stone 0.031 0.062 -0.031

The gain is concentrated in the early wood/stone tech tree. make_stone_pickaxe is the one clear regression, and it is odd next to make_stone_sword improving — at 32 agents each achievement is 32 Bernoulli trials, so a 0.094 difference is 3 agents and well inside sampling noise.

Caveat carried from the section above

These numbers use --mode inference's sample_plan_inpainting sampler and are not comparable to results.json's pretrained_score scale.

Downloads last month
42