diff --git a/Justfile b/Justfile index 3d52c6b..48f4d06 100644 --- a/Justfile +++ b/Justfile @@ -116,7 +116,19 @@ align-premises: # Full retrain from provisioned corpora. Honours ECHIDNA_MAX_PROOF_STATES # (0 = unlimited), ECHIDNA_NUM_EPOCHS, ECHIDNA_NUM_NEGATIVES. -retrain: +# +# Depends on align-premises: without that step the dataloader joins +# premise.proof_id against proof_state.id and fails at ~0% match because +# merge_corpus.jl rewrites proof_state ids to fresh sequential counters +# while the premise files keep the original per-extractor ids. Running +# align-premises is idempotent and cheap (seconds on a 2M premise set). +retrain: align-premises + julia --project=src/julia src/julia/run_training.jl + +# Retrain without the align-premises prerequisite — use only when you +# know premises_COMPLETE.jsonl is already aligned (e.g. you just ran +# just align-premises manually and want to iterate on model code). +retrain-skip-align: julia --project=src/julia src/julia/run_training.jl # End-to-end pipeline: provision → extract → merge → align → retrain. diff --git a/crates/typed_wasm/Cargo.toml b/crates/typed_wasm/Cargo.toml index c27f2d2..fe23a6f 100644 --- a/crates/typed_wasm/Cargo.toml +++ b/crates/typed_wasm/Cargo.toml @@ -11,7 +11,3 @@ repository = "https://github.com/hyperpolymath/echidna" [dependencies] anyhow = "1" -async-trait = "0.1" -tokio = { version = "1", features = ["full"] } -serde = { version = "1", features = ["derive"] } -serde_json = "1"