Upload v5/imagination_model_v5.pt with huggingface_hub

Upload v5/imagination_model_v5.pt with huggingface_hub

Upload v5/bc_stage.pt with huggingface_hub

Upload v5/bc_stage.pt with huggingface_hub

v5.3 fix: Actor subclasses nn.Module (register_buffer crash that killed run 6aafa538)

Fix Actor: subclass nn.Module (register_buffer/parameters/to/state_dict now work)

Upload fix3/imagination_model.pt with huggingface_hub

v5.3: BC on oracle-verified labels only (omask) — validated end-to-end CPU smoke EXIT=0, BC CE 0.0014 on clean labels

v4: scale-up (same method): dm 384, 6 layers, rl 12000 steps, n_demo 600 x 3 tasks, cls 2000

v4: scale-up (dm 384, 6 layers, 600 demos/task, 12000 RL steps) - same method, more capacity and data

Upload v5/imagination_model_v5.pt with huggingface_hub

Upload v5/bc_stage.pt with huggingface_hub

v5.2 full validated script (overwrite truncated 4.4KB partial from commit 6ec8e27)

v5.2: spatial policy head, 6x6 action grid aligned with FSQ token grid, advantage normalization removed (entropy-collapse root cause), grad clip 1.0, standard entropy bonus; smoke-validated end-to-end on CPU

Upload v5/bc_stage.pt with huggingface_hub

v5.1: fix classifier device crash, real-vs-imagined label mixing, quadratic entropy floor, BC-stage checkpoint push

v5 verified script (sandbox EXIT_0, run9) + trackio logging

Upload v3b/imagination_model_v3.pt with huggingface_hub

v5: DINOv2+SmolLM2 actor features, BC-then-RL, pessimistic reward ensemble, next-frame CE fix

Upload v4/imagination_model.pt with huggingface_hub

v5 real content: rich actor inputs (DINOv2 + SmolLM2 text emb), text-aware WM, BC-then-RL; CPU smoke passed

Upload fix2/imagination_model.pt with huggingface_hub

v5: actor-representation fix (frozen DINOv2 patch feats + frozen SmolLM2 text embedding into the actor), text-aware WM, BC-then-RL; CPU smoke passed all phases

Upload fix2/imagination_model.pt with huggingface_hub

v3: entropy-target control (target 4.5, penalize below), rl_steps 4000 (timeout-driven), ent_coef 0.05

v4 real content: v2 pipeline + BC phase D0 (oracle-only demos, lr 1e-3) before RL; ent_coef 0.05; CPU-smoke verified

v4: demos-first behavior cloning (phase D0, oracle-only, lr 1e-3) before RL; ent_coef 0.05; CPU-smoke verified

Upload fix2/imagination_model.pt with huggingface_hub

v3 smoke-verified: fix imagine_step 1-D hidden handling (sandbox CPU smoke passed all phases)

fix: set pad_token=eos_token; use tokenizer attention_mask instead of ids!=0 heuristic

v3: frozen DINOv2 encoder -> 36x6 FSQ -> SmolLM2-135M world model with real text instructions

v3: ent_coef 0.01 -> 0.05 to stop policy entropy collapse (v2 diagnosis: WM generalizes fine, val CE == train CE; actor entropy hit 0.0 by RL step 2666)

Upload fix1b/imagination_model.pt with huggingface_hub

fix: set pad_token = eos_token on SmolLM2 tokenizer (has no pad token)

v3 backbone version: frozen DINOv2-small encoder + 36-token 6-dim FSQ (70 bytes/frame), SmolLM2-135M world model with real text instructions, KV-cache imagination

v2.1: ent_coef 0.01 -> 0.02 (entropy collapsed to 0 under advantage normalization in canceled fix1 run)

v2: anti-memorization (90/10 val split + WM early stopping), 2-head reward ensemble with pessimistic min in imagination, Lc=3 contexts, advantage normalization, 3-task demos

Upload real/imagination_model.pt with huggingface_hub

Upload smoke/imagination_model.pt with huggingface_hub

Add verified MiniWoB++ imagination-model training script (smoke-tested end to end in sandbox)

initial commit