repo stringclasses 20
values | path stringlengths 6 94 | lang stringclasses 5
values | n_chars int64 81 200k | sha256 stringlengths 64 64 | content stringlengths 81 200k |
|---|---|---|---|---|---|
eren23/crucible-community-tap | objectives/sigreg/sigreg.py | py | 5,955 | 07333b7a8b73a66592e23ab73696326b3fbcb7fddda5c4b8384d07929313ff77 | """SIGReg: Sketched-Isotropic-Gaussian Regularizer.
Prevents representation collapse in JEPA models by enforcing that latent
embeddings follow an isotropic Gaussian N(0, I), using the Cramér-Wold
theorem. If all 1D marginal projections are Gaussian, the full distribution
is Gaussian.
Computes M random projections of ... |
eren23/crucible-community-tap | objectives/distillation/distillation.py | py | 4,913 | 1c901570b2e3c448c725abe1559eaa38d9c9b249eab7d2aa9eb30395bae6c8ac | """Knowledge distillation objective for generic_backend.
Wraps a teacher model and computes the combined loss:
L = alpha * T^2 * KL(student_soft || teacher_soft) + (1 - alpha) * task_loss
Env vars:
DISTILL_TEACHER_PATH: Path to teacher state dict (required)
DISTILL_TEACHER_FAMILY: Model family for teach... |
eren23/crucible-community-tap | objectives/state_prediction/state_prediction.py | py | 2,997 | 7ed567a8a795acb3740ac484a77ed050796e84d05f589b18a1fd04a98af39005 | """State prediction objective -- JEPA-style latent state prediction with variance regularization.
Designed for world models that produce predicted and target embeddings
in latent space. Combines MSE prediction loss with VICReg-style variance
hinge regularization to prevent representation collapse.
Expects:
predic... |
eren23/crucible-community-tap | optimizers/lion/lion.py | py | 2,587 | f0a22f274e09cabcfc85b2b47945129c94ef278e3ba23d8a9729bb6193541a08 | """Lion optimizer — EvoLved Sign Momentum (Chen et al., 2023).
Sign-based update with decoupled weight decay. Memory-efficient
alternative to Adam: only stores momentum (no second moment),
halving optimizer state memory.
Usage:
EMBED_OPTIMIZER=lion MATRIX_OPTIMIZER=lion ...
"""
import torch
from crucible.training... |
eren23/crucible-community-tap | data_adapters/trajectory_hdf5/trajectory_hdf5.py | py | 8,407 | 459a3850b7b8e93d1537fc80c9dc62d8ad8a515550bb03f694971a5d9e5a1c59 | """HDF5 trajectory data adapter for world model training.
Loads trajectory datasets with frames, actions, and optional goals from
HDF5 files. Returns sub-trajectory slices for JEPA-style training.
Expected HDF5 layout::
/trajectories/
/0/
frames: [T, H, W, C] uint8 or [T, C, H, W] float32
... |
eren23/crucible-community-tap | data_adapters/code_state/trajectory_adapter.py | py | 9,303 | 3f21e5fe3d746abc394ff798de1fecfedd09b6ced87b34a13d3af24dfc788f05 | """Trajectory-aware data adapter for multi-step world model training.
Loads trajectory-structured HDF5 (produced by TrajectoryCollector) and serves
multi-step (state, action, next_state) chains. Falls back to single-step
random sampling when trajectory metadata is absent.
Each batch returns a window of ``window_len`... |
eren23/crucible-community-tap | data_adapters/code_state/code_state.py | py | 8,662 | 85610983aa93abc910fa351c4c5b83b21921b6dd5cd7570ec66642d85ed7b46b | """Code state data adapter -- loads git edit HDF5 for world model training.
Loads HDF5 files produced by the ``git_edit`` or ``commitpack`` collectors
and serves (state, action, next_state) tuples for world model training.
Each transition is a localized code change:
- **states**: AST token IDs of file content *before... |
eren23/crucible-community-tap | examples/diff_xyz/diff_xyz_train_qwen.py | py | 13,853 | 200f63645057cfeb429ace21fba071859896a5f1f1527d2ff9d6a21755ce9d08 | # ---
# jupyter:
# jupytext:
# text_representation:
# extension: .py
# format_name: percent
# format_version: '1.3'
# jupytext_version: 1.16.0
# kernelspec:
# display_name: Python 3
# language: python
# name: python3
# ---
# %% [markdown]
# # Fine-tune Qwen3-Coder on Diff-XY... |
eren23/crucible-community-tap | examples/diff_xyz/diff_xyz_colab.py | py | 6,237 | ba7405b08981dfad1707a4e2253e9d46651196844a04c5abe9b9e9a00293ee7a | # ---
# jupyter:
# jupytext:
# text_representation:
# extension: .py
# format_name: percent
# format_version: '1.3'
# jupytext_version: 1.16.0
# kernelspec:
# display_name: Python 3
# language: python
# name: python3
# ---
# %% [markdown]
# # Diff-XYZ Colab Smoke
#
# Runs th... |
eren23/crucible-community-tap | launchers/_launcher_common.py | py | 11,797 | 8e15e4c048b7bc37377b75fb425683a5b61d893e94a5c4555cc1efcd0cb4aec3 | """Shared launcher utilities for LE-WM family launchers.
Provides env-var parsing helpers, config loading, Hydra override application,
and metadata writing that were copy-pasted across lewm_upstream,
hybrid_lewm_upstream, and elastic_lewm_upstream.
"""
from __future__ import annotations
import json
import os
from pat... |
eren23/crucible-community-tap | launchers/lewm_upstream/precompute_pusht.py | py | 9,513 | c7033ff2ca96b7af8fc3ba9fef04b61f8d2d8f0083880b03987b99f81b17238f | """Precompute PushT HDF5 dataset into a HuggingFace dataset with transforms baked in.
Runs inside the upstream ``lucas-maes/le-wm`` workspace (same as the launcher).
Loads the HDF5 data via ``stable_worldmodel``, applies the same transforms
(image resize + column normalization) that the launcher applies at training
ti... |
eren23/crucible-community-tap | launchers/lewm_upstream/lewm_upstream.py | py | 4,997 | 1464bc4995955684324e5d9c49489f56ea9c22fb17c18298597eec108a016e97 | """Reusable LE-WM launcher bundle for Crucible external projects.
This launcher runs inside a cloned upstream ``lucas-maes/le-wm`` workspace.
It keeps the project spec thin while making the experiment entrypoint
shareable through local plugins, installed hub packages, or tap clones.
"""
from __future__ import annotati... |
eren23/crucible-community-tap | launchers/hybrid_lewm_upstream/hybrid_lewm_upstream.py | py | 23,931 | bf43ac1271ceed138af3d869e059ba516daabc308e8a22263fd73276d9b2062c | """Hybrid LE-WM upstream launcher for Crucible external projects.
Fork of lewm_upstream launcher that replaces the upstream HF ViT encoder with
a HybridViTEncoder (configurable mix of softmax + linear attention blocks).
Runs inside a cloned upstream ``lucas-maes/le-wm`` workspace. All hybrid
encoder classes are defi... |
eren23/crucible-community-tap | launchers/elastic_lewm_upstream/elastic_lewm_upstream.py | py | 27,021 | 7b7b571d688852ea7a41119057cf91841c50c98875a9ed67430e60741635d5f2 | """Elastic LE-WM launcher bundle for Crucible external projects.
Fork of lewm_upstream.py with elastic compute routing. This launcher runs
inside a cloned upstream ``lucas-maes/le-wm`` workspace, replacing the
encoder and predictor with elastic versions and adding the DifficultyRouter.
All elastic model components ar... |
eren23/crucible-community-tap | launchers/code_wm/train_code_wm.py | py | 52,203 | 8f833fee408646f8025a35eea104e89236bfe98577fd01b21da9e0924af71fbf | #!/usr/bin/env python3
"""Standalone Code World Model training script for RunPod / GPU execution.
Supports two data modes:
- **Single-step** (default): classic (before, action, after) pairs from flat HDF5
- **Trajectory**: multi-step windows from trajectory-structured HDF5
(auto-detected via /metadata.has_traj... |
eren23/crucible-community-tap | launchers/code_wm/pretrain_encoder.py | py | 11,990 | 927c2ab61f6e5428246a9e9317d8734f17e8773db8b6315681ab39c7cf0c1360 | #!/usr/bin/env python3
"""Pre-train CodeWM encoder with supervised objectives (Phase 12, Direction B).
Trains the state encoder to predict edit properties from the before-state
encoding. This forces the encoder to learn representations that distinguish
different code contexts — preventing the rank collapse seen with J... |
eren23/crucible-community-tap | launchers/kernel_wm/train_kernel_wm.py | py | 18,980 | 448934e932f0434de9c4c7548128d2df8a2357bfc9c0115d630c85f76a240989 | #!/usr/bin/env python3
"""KernelWM training script — trains CodeWM on GPU kernel migration pairs.
Uses the same CodeWorldModel architecture with adjusted vocab (400) and
action_dim (12). Adds example table logging: during validation, decodes
predicted kernel configs back to readable form and saves comparison tables.
... |
eren23/crucible-community-tap | launchers/code_deltatok/train_deltatok.py | py | 13,348 | 8b6162378256a03a4408ce2172565596ec980c61c969a3cc5b01e5bd1b7d254a | #!/usr/bin/env python3
"""CodeDeltaTok training script.
Trains a DeltaTok-inspired delta tokenizer on pre-computed frozen backbone
features. No backbone inference during training -- just loads feature vectors
from HDF5 and learns to compress deltas.
Env vars:
CDT_HDF5_PATH: Path to pre-computed features HDF5
... |
eren23/crucible-community-tap | launchers/diff_xyz/sft_smoke.py | py | 15,849 | 26ef1fb342d115d7a5a09032f59ca27005b308fdc69b6eae509669631c8cdac5 | """SFT-driven Diff-XYZ smoke runner.
Loads a small Qwen base in 4-bit, attaches LoRA, fine-tunes for a handful of
steps on a CommitPackFT subset (deduped against Diff-XYZ test), then runs
post-training eval on a small Diff-XYZ slice. Written to fit a 24GB GPU
(RTX 4090 / 3090 / A5000) and finish in ~5 minutes.
Env va... |
eren23/crucible-community-tap | launchers/diff_xyz/api_smoke.py | py | 2,908 | 2f1d9793438db9e197db49cfe62e961353a1cdfe67e23ed63601d6b97959d36d | """API-driven Diff-XYZ smoke runner.
Invokes `evaluation.diff_xyz.harness.run_benchmark` against OpenAI / Anthropic /
Google APIs. Used to validate the harness on a small sample before committing
to a full sweep.
Env vars (all optional with sane defaults):
DIFFXYZ_MODEL model spec, default 'openai:gpt-4... |
eren23/crucible-community-tap | launchers/diff_xyz/eval_only.py | py | 8,420 | e1e2d150067d2c0b585dd3f1976b4aba01233e747c9729565310f12bc1a1d0d6 | """Zero-shot Diff-XYZ eval runner — local HF model, no SFT.
Loads a base HF model in 4-bit and runs the Diff-XYZ harness against it.
The point: anchor our pipeline against the paper's no-SFT baselines (e.g.
Table 5: Qwen2.5-Coder-7B Diff-Gen search-replace EM=0.68). Without that
anchor we can't tell whether a fine-tun... |
eren23/crucible-community-tap | evaluation/rescore_wandb_history.py | py | 4,430 | 595b9afdc89db69da46afca228fb869cbf4c0e8ab98b4e5a6e16dc0530c5477a | #!/usr/bin/env python3
"""Rescore wandb runs with mean-of-last-N instead of peak-only.
Fetches metric history for every run in ``eren23/crucible-code-wm`` and
reports, per run:
- peak val_delta_cos (the number we used to report)
- mean ± std of val_delta_cos over the last N eval steps (default N=5)
- total steps run
... |
eren23/crucible-community-tap | evaluation/baselines.py | py | 13,018 | e7bf0719d2d647d5b2bd4fb4273d15f2460e689530e9ab690323c3e37261f009 | #!/usr/bin/env python3
"""Baselines for the Code WM paper.
Answers the question: "Is CodeWM's 92-95% edit classification and 0.14
delta_cos actually impressive, or does a trivial method get close?"
Three baselines that work directly on the existing HDF5 data (no
retokenization needed):
1. Bag-of-AST-tokens: cosine s... |
eren23/crucible-community-tap | evaluation/code_wm/rollout_eval.py | py | 14,101 | 33d7f7190ead48107d2b4fde4f7da81b189eb9a3e6f5f83a3a0dced2eb417982 | #!/usr/bin/env python3
"""Multi-step rollout evaluation for Code World Model.
Two subcommands:
- **drift**: Phase 1 sanity check. Sample N transitions from the training h5,
roll out the predictor with a fixed action for several steps, report
magnitude + cosine drift. No ground truth required.
- **trajectory**: P... |
eren23/crucible-community-tap | evaluation/code_wm/latent_planning.py | py | 8,334 | d90130bfd5bf24d5e9f7c2e7593546f2e85b5efd2daaf8351dbcefa30c6ba532 | #!/usr/bin/env python3
"""Latent gradient planning eval for CodeWM.
Per the temporal-straightening / planning prior art (arxiv 2603.12231):
given z_before = encoder(state_t) and z_target = target_encoder(state_{t+1}),
can we recover the latent action via gradient descent in latent space?
z_a* = argmin_z_a || p... |
eren23/crucible-community-tap | evaluation/code_wm/modern_baselines_compare.py | py | 11,280 | 9ab5def09f14b1e36cc25718a2c96f88b1f946e98192dd97828f0f2debdf2913 | #!/usr/bin/env python3
"""Modern code embedding baselines vs CodeWM on CodeSearchNet retrieval.
Closes the "stale baselines" gap (CodeBERT 2020 only) from the Phase 5C
benchmark audit by running 2024-2026 code embedding models on the same
CodeSearchNet code-to-code retrieval protocol that ``codesearchnet_eval.py``
use... |
eren23/crucible-community-tap | evaluation/code_wm/delta_magnitude_analysis.py | py | 10,689 | b8c8dcd72e936d48bd98dd15a286437f966943a4d14cafc93168754973c557c4 | #!/usr/bin/env python3
"""Phase 0.4 — Delta magnitude distribution analysis.
Computes ||z_{t+1} - z_t|| / ||z_t|| across datasets to quantify the
signal-to-noise ratio for delta prediction. Designed to compare CodeWM
(Python, expected low signal) vs KernelWM (CUDA kernels, expected high
signal) to confirm the domain h... |
eren23/crucible-community-tap | evaluation/code_wm/_shared.py | py | 4,980 | 42c247a68a956e97cd3734b3083801af6773e0a1ce5695423fba5745b6a582f4 | """Shared CodeWM eval helpers.
Canonical implementations for the duplicated functions that used to live
in every single `evaluation/code_wm/*.py` script. Pulled out into one
module so a bug fix (e.g. the 2026-04-10 ``WM_POOL_MODE=cls``/
``strict=False`` silent-drop bug) only needs to be made once.
Public API:
- ``lo... |
eren23/crucible-community-tap | evaluation/code_wm/semantic_eval.py | py | 10,880 | 373774e0c04943a9cd3fa96f491684466fb9ac4e3eba1518eaac1ca9b1911a43 | #!/usr/bin/env python3
"""Semantic downstream tests for Code World Model.
Tests whether the learned latent geometry is actually useful, beyond
val/cosine_sim. Four probes:
1. Edit Retrieval: For each query edit, find k-NN in delta space.
Measure action-vector similarity between query and retrieved neighbors.
2. k... |
eren23/crucible-community-tap | evaluation/code_wm/null_models_rollout.py | py | 7,529 | aa34834e83dd33a3c01a03531957bc197b578688c0407f9629b1d9c87bca8a06 | """Delta-space null models for the multi-step compositional prediction claim.
The paper reports the per-step metric s_k = cos(pred_delta_k, true_delta_k)
where pred_delta_k is the predictor's latent displacement between rollout
step k-1 and step k, and true_delta_k is the ground-truth target-encoder
displacement. The... |
eren23/crucible-community-tap | evaluation/code_wm/isotropy_analysis.py | py | 9,124 | 4857a05911cdb60f5edd8eff74848d4554e16373a460a0d13e0cf83811798713 | #!/usr/bin/env python3
"""Phase 0.3 — Isotropy analysis of encoder latent space.
Measures how close the frozen encoder's representation space is to an
isotropic Gaussian (the optimal geometry for delta prediction per
LeJEPA/SIGReg theory). An anisotropic space means some delta directions
are privileged over others, ma... |
eren23/crucible-community-tap | evaluation/code_wm/codesearchnet_eval.py | py | 10,262 | c7cdc57b49171eb1cece15567093ffcd9e2eb7f0eff5244e2deddc95cade6b55 | #!/usr/bin/env python3
"""External-benchmark retrieval eval on CodeSearchNet Python test split.
Protocol (code-to-code retrieval on a diverse external corpus):
1. Load ``code_search_net`` Python test split (22,176 functions from open-source
Python projects, each annotated with its origin ``repository_name``).
2. S... |
eren23/crucible-community-tap | evaluation/code_wm/delta_norm_report.py | py | 9,617 | bb07afdf9f7807a362728a5faef4ca323bb7b725f8ee4184d1c65552e15126bb | #!/usr/bin/env python3
"""Delta magnitude / norm reporting for CodeWM.
Closes the "cosine alone can hide whether the model is predicting a tiny
vector" concern from the 2026-04-11 reviewer-response pass. For each
checkpoint we sample N single-step transitions (and optionally longer
trajectories for s2/s3) from the tra... |
eren23/crucible-community-tap | evaluation/code_wm/geometry_probes.py | py | 11,966 | 94f63912ab9190a9fc91a64abcb88ac50bdb83c16a6af321db5c4b7616ee51f6 | #!/usr/bin/env python3
"""Phase 0.1 — Layer-wise linear probes for CodeWM encoder.
Probes intermediate representations after each loop iteration of the
weight-shared LoopedTransformerBlock to answer:
1. Which loop iterations contain transition information?
2. Can linear probes predict the next-state delta directio... |
eren23/crucible-community-tap | evaluation/code_wm/cross_repo_eval.py | py | 14,177 | c0d942043179b41a86c8ff05c4e623a321e0c6f1d277031680720f86dc851e99 | #!/usr/bin/env python3
"""Cross-repository retrieval eval on 20 held-out Python repos.
Closes the "cross-repo generalization never measured" gap from the Phase 5C
audit. The model was trained on CommitPackFT Python (IID val split). Here we
test it on fresh git histories of 20 well-known Python libraries that are
NOT i... |
eren23/crucible-community-tap | evaluation/code_wm/curvature_analysis.py | py | 10,049 | 8032be4dd09d4420777e682a083040eaa24e80ad91bae248d4ee2e2f01bfa99a | #!/usr/bin/env python3
"""Phase 0.2 — Trajectory curvature measurement in latent space.
For sequences of 3+ consecutive commits, computes curvature of the
trajectory in encoder latent space:
kappa = ||delta_{t} - delta_{t-1}|| / ||delta_{t-1}||
where delta_t = z_{t+1} - z_t.
Low curvature (straight trajectories... |
eren23/crucible-community-tap | evaluation/code_wm/commit_chronicle_eval.py | py | 18,689 | c3462bb61bfc848aa8fa009d8a7525b3ae69b1f380c1c8e110535bff38f4e30b | #!/usr/bin/env python3
"""Evaluate CodeWM vs CodeBERT vs BoW on CommitChronicle subset_llm.
JetBrains-Research/commit-chronicle subset_llm has 4,030 commits with per-file
diffs and reference commit messages. The retrieval task: given a (before, after)
code edit, retrieve the most similar edit by delta-NN.
Relevance c... |
eren23/crucible-community-tap | evaluation/code_wm/full_val_knn.py | py | 8,808 | bcb7f1f9a1cbd76b3687506db7589dc94ea06b4d2aa76cf9146b157c9bf21cf6 | #!/usr/bin/env python3
"""Full-val KNN + eff_rank evaluation on a CodeWM checkpoint.
Addresses the B1 audit finding: training-time val/knn metrics run on a
single 128-sample batch and are noisy. This script runs the same
metrics on the FULL val set (typically 5000 samples) for a defensible
paper number.
Computes:
-... |
eren23/crucible-community-tap | evaluation/code_wm/honest_retrieval.py | py | 14,280 | 19c1721bda4c5e44ae5f6d12997fedf107da864bdd9d096c8bd137b4037ca0f9 | #!/usr/bin/env python3
"""Honest retrieval evaluation that doesn't get fooled by dataset imbalance.
The existing semantic_eval.py reports MRR/Recall@k with "same edit_type" as the
relevance criterion. But CommitPack is 98.8% MODIFY, so this metric is trivially
saturated — random retrieval scores ~0.99.
This script re... |
eren23/crucible-community-tap | evaluation/code_wm/delta_diagnosis.py | py | 8,192 | 2ab9f0a8757df0642ff912e8b5fd8f7cd9daecafdf4db4774621c63f608f988a | #!/usr/bin/env python3
"""DeltaCodeWM Generalization Diagnosis.
Tests why in-batch training metrics (dcos=0.957, lift=+0.128) don't transfer
to independent holdout (pred_vs_diff cosine=0.05).
Usage:
python delta_diagnosis.py --checkpoint checkpoints_delta/kernel_wm_best.pt --data data/commitpackft_with_diffs.h5
"... |
eren23/crucible-community-tap | evaluation/code_wm/eval_code_wm.py | py | 9,574 | 7fedb67238d814c3f6c7e400a6043e399c876e04d67598bfffd8c3274e3698c1 | #!/usr/bin/env python3
"""Evaluate a Code World Model on held-out git edit transitions.
Computes:
- MSE in latent space (pred vs target)
- Cosine similarity between predicted and target embeddings
- Nearest-neighbor accuracy (does nearest neighbor of pred = actual target?)
Compares against baselines:
... |
eren23/crucible-community-tap | evaluation/code_wm/codebert_compare.py | py | 12,791 | 368183f9303970d79ac07aa6bd8c4b4533aa0687a3f704d3fd8698b847da2a7f | #!/usr/bin/env python3
"""Head-to-head retrieval comparison: CodeWM vs CodeBERT vs BoW.
Downloads CodeBERT (microsoft/codebert-base) and a fresh CommitPackFT sample,
encodes the same edits with both models, runs the same delta-NN retrieval
pipeline, and compares MRR/Recall@k under multiple relevance criteria.
Why fre... |
eren23/crucible-community-tap | evaluation/code_wm/cross_repo_modern_compare.py | py | 18,559 | 93a2cac123b1cd255915e4b05fb3d2cc47476eca7b372bb3a8e59fb53d48bfb8 | #!/usr/bin/env python3
"""Cross-repo retrieval eval with modern baselines (CodeBERT, CodeT5+, jina).
Extends ``cross_repo_eval.py`` to also embed the (before, after) pairs with
modern HF code embedders, not just CodeWM + BoW. Uses the same 20-repo basket
and the same leave-one-repo-out protocol so results are directly... |
eren23/crucible-community-tap | evaluation/kernel_wm/cross_arch_transfer.py | py | 13,372 | 7a92522c7d83e175d31880e4a881ddd4c49003addf5c326936763b47f2d59b4a | #!/usr/bin/env python3
"""KernelWM Cross-Architecture Transfer Evaluation.
The moonshot test: train on SM80->SM90 + SM90->SM100, test zero-shot
on SM80->SM100. Compares model against copy-last, random, and NN baselines.
Usage:
python cross_arch_transfer.py \
--checkpoint kernel_wm_best.pt \
--trai... |
eren23/crucible-community-tap | evaluation/code_deltatok/eval_diffxyz.py | py | 7,072 | a3e5cc2b52e7bd82ebc309707a45c6c5087cb0b5dbdf988ff040cdfcc4241431 | #!/usr/bin/env python3
"""CodeDeltaTok evaluation on Diff-XYZ public benchmark.
Downloads Diff-XYZ test set from HuggingFace, encodes with UniXcoder,
runs CodeDeltaTok, and evaluates cross-modal retrieval.
This anchors the paper's claims to a public benchmark.
Usage:
python eval_diffxyz.py \
--checkpoint... |
eren23/crucible-community-tap | evaluation/code_deltatok/eval_efficiency.py | py | 6,170 | 7f1c24814a69a43321c03761ba198c6590868e6f2e0872e9433a3c1f0c923ba7 | #!/usr/bin/env python3
"""CodeDeltaTok efficiency benchmark.
Measures the systems advantage of 1-token delta representation:
- Encoding latency (ms/change)
- Storage per change (bytes)
- ANN retrieval latency at scale
- Comparison to full-feature storage
Usage:
python eval_efficiency.py \
--featur... |
eren23/crucible-community-tap | evaluation/code_deltatok/eval_deltatok.py | py | 11,431 | adc262357ca2d83a0f04e22866d2f244f744feeb8f30e45346bb4e022dd46cdb | #!/usr/bin/env python3
"""CodeDeltaTok benchmark — proper downstream evaluation.
Three evaluation tasks:
Task 1: Cross-modal retrieval
Query = delta representation (from method under test)
Gallery = actual after-state features (frozen UniXcoder)
Ground truth: query[i] should match gallery[i] (same code change)
... |
eren23/crucible-community-tap | evaluation/code_deltatok/eval_hard_negatives.py | py | 7,919 | 64f3b4b51f636bc32ba1876e009612727887c0b7622773a668752c60281ae6f2 | #!/usr/bin/env python3
"""CodeDeltaTok hard-negative evaluation.
Tests whether delta tokens can discriminate between changes when the
starting code is similar. This is the key reviewer defense: "MRR 0.997
with random negatives is too easy."
Hard negative construction:
For each query, gallery contains K hard negativ... |
eren23/crucible-community-tap | evaluation/diff_xyz/metrics.py | py | 5,915 | c51e4b801b7e3428d7c6219a42ff5abf8002ae1109f36833911730cb4fd8d048 | """Diff-XYZ metrics (paper §3.1 + §3.2 verbatim).
- Stripped EM: whitespace-only lines removed, then exact string equality
- Stripped IoU: |unique lines A ∩ B| / |unique lines A ∪ B| after whitespace-strip
- F1 on added lines: F1 between {added_lines in ref} and {added_lines in pred}
- F1 on deleted lines: analogous
-... |
eren23/crucible-community-tap | evaluation/diff_xyz/harness.py | py | 10,077 | 402bbd90fe9811fc1e7a8f846ed7f8cbcf26eb08867199ba2fe7e0802d77a00b | """Diff-XYZ evaluation harness — the CLI that wires every piece together.
Usage:
python evaluation.diff_xyz/harness.py \\
--model anthropic:claude-sonnet-4-6 \\
--task apply \\
--format udiff \\
--system-prompt format \\
--limit 20 \\
--out /tmp/smoke.json
Output JS... |
eren23/crucible-community-tap | evaluation/diff_xyz/models.py | py | 15,653 | ae31ab62bfb02fdb764abeb23917a70f742a28c586fba0f74bee4417e8df61ef | """Model adapters for the Diff-XYZ harness.
Protocol:
generate(system_prompt: str, user_prompt: str, *, max_tokens: int = 4096,
temperature: float = 0.0) -> str
Backends:
- anthropic:<model_id> via `anthropic` SDK
- openai:<model_id> via `openai` SDK
- google:<model_id> via `goog... |
eren23/crucible-community-tap | evaluation/diff_xyz/__init__.py | py | 859 | 7a124980e36496993d7aa4fcafe449ef80400579be913dbb81a43608eed16bb4 | """Diff-XYZ eval harness (arXiv 2510.12487).
Generative benchmark for LLM code-diff understanding:
- Apply: old_code + diff -> new_code
- Anti-Apply: new_code + diff -> old_code
- Diff-Gen: old_code + new_code -> diff
Public API:
from evaluation.diff_xyz import compute_metrics, parse_diff, apply_d... |
eren23/crucible-community-tap | evaluation/diff_xyz/prompts.py | py | 6,822 | 83accc0ff210fb0fa5298e59ea18848bbdeee7272e8172e72f999de76ed315a6 | """System + user prompt templates for Diff-XYZ tasks.
Paper Appendix A describes the prompts but the arxiv HTML export renders them
as figure placeholders without figure content. The templates below are a
best-effort reproduction matching the paper's description:
- `w/o format` system prompt: literally "You are a h... |
eren23/crucible-community-tap | evaluation/diff_xyz/dataset.py | py | 3,214 | deef766bc538fdee0073749b98c78afb4bc5a0870e7f36e3ffe8f882a9645256 | """Load the Diff-XYZ test set from HuggingFace.
Dataset: `JetBrains-Research/diff-xyz`. Single `test` split, 1000 rows,
columns per dataset card: repo, commit, path, lang, license, message,
old_code, new_code, n_added, n_removed, n_hunks, change_kind,
udiff, udiff-h, udiff-l, search-replace.
`load_samples(limit, lang... |
eren23/crucible-community-tap | evaluation/diff_xyz/formats.py | py | 11,368 | 5e66c8c9cf84764234a5e6231bc1c592b843da57f846524b972e9491a699f5f0 | """Diff-XYZ format parsers + appliers.
Formats (per paper §5):
- udiff: standard unified diff with `@@ -a,b +c,d @@` hunk headers
- udiff-h: unified diff with relaxed `@@...@@` headers (no line nums)
- udiff-l: verbose `ADD`/`DEL`/`CON` markers instead of `+`/`-`/` `
- search-replace: li... |
eren23/crucible-community-tap | evaluation/diff_xyz/tests/test_diff_xyz_metrics.py | py | 8,773 | ade8f65d8f6c0753a2f870bc5b81366d94ffdb903ad06aec5ec8ce6040640650 | """Tests for evaluation.diff_xyz/metrics.py and formats.py."""
from __future__ import annotations
import pytest
from evaluation.diff_xyz.formats import (
ApplyError,
ParseError,
apply_diff,
diff_added_lines,
diff_deleted_lines,
parse_diff,
)
from evaluation.diff_xyz.metrics import (
comput... |
eren23/crucible-community-tap | evaluation/diff_xyz/tests/test_diff_xyz_harness.py | py | 9,507 | c31d6d0a99e3c4358078ed3965aec2fb73cdc1903fac893b70bc906911ebbba3 | """Tests for evaluation.diff_xyz/harness.py end-to-end flow with a dummy backend.
Full API-backed runs live in a separate integration test (not run in CI).
"""
from __future__ import annotations
import json
import sys
from pathlib import Path
from unittest.mock import MagicMock, patch
import pytest
from evaluation.... |
eren23/crucible-community-tap | evaluation/diff_xyz/tests/test_diff_xyz_models.py | py | 4,739 | 1f489cfa20371ec8c1fe54c6a137254d646cd8cd46ebdbc0fb6d48bc11d27b4a | """Tests for evaluation.diff_xyz/models.py (backend resolution + dummy only).
API-backed adapters are covered by the end-to-end smoke test that runs against
live keys, not here.
"""
from __future__ import annotations
import pytest
from evaluation.diff_xyz.models import (
AnthropicBackend,
DummyBackend,
G... |
eren23/crucible-community-tap | evaluation/diff_xyz/tests/test_diff_xyz_prompts.py | py | 6,248 | dd69761c83ee8e469441391600a9916c8d86adbf361266907c135e57f3ae52bd | """Tests for evaluation.diff_xyz/prompts.py and dataset.py (offline portions)."""
from __future__ import annotations
import pytest
from evaluation.diff_xyz.dataset import DiffXYZSample, _row_to_sample
from evaluation.diff_xyz.prompts import (
GENERIC_SYSTEM,
SYSTEM_PROMPT_MODES,
TASKS,
strip_markdown_... |
eren23/coder-interp-tap | callbacks/wandb_periodic_validation/wandb_periodic_validation.py | py | 2,938 | 6495f34ea11549f79a66503dddaafeb2bc1aca40e28d4b88e0695267d8c89e07 | """Skeleton periodic-validation callback for W&B logging.
Drops into a Crucible training loop. Every `eval_interval` steps it runs a
validation pass and logs scalar metrics and a sample table to W&B. Every
`checkpoint_interval` steps it saves the model state.
This is a SKELETON. Wire `run_validation` to your actual e... |
eren23/coder-interp-tap | findings/2026-05-10-cross-author/make_plots.py | py | 6,358 | c9c02706047629e333ec688e000e78a300b2e5da4c46561cc5ccf774c55e521d | """Generate visualizations from the v1/v2/v3 cross-author delta study.
Pulls live W&B data for the 9 completed runs and produces 4 PNGs in this
directory. Real numbers, no synthesis — change the V1/V2/V3 dicts to point
to a new round of runs to regenerate.
Run:
cd findings/2026-05-10-cross-author
python3 make... |
eren23/coder-interp-tap | launchers/nla_coder/av_train.py | py | 784 | 94d133d7f9564993713f99629b25c67214ab0484c765c2da07cd17f2c68dca2f | """Phase-2.2 AV training launcher for nla_qwen2_5_coder_1_5b.
SCAFFOLD ONLY. Real implementation goes here when Phase 2.1 SAE
training is producing usable feature activations.
Intended pipeline:
1. Load saved Phase-2.1 SAE checkpoint.
2. For each of d_sae features, generate synthetic NL description via
Claud... |
eren23/coder-interp-tap | launchers/nla_coder/sae_train.py | py | 807 | b5fcac841e37ea045f9c6826535934649b95ec7bd71cef777ce9480d2478be2e | """Phase-2.1 SAE training launcher for nla_qwen2_5_coder_1_5b.
SCAFFOLD ONLY. Real implementation goes here when Phase 0 smoke is green.
Intended pipeline:
1. Stream ~100M code tokens from CommitPackFT (Python + TS subsets).
2. Forward through Qwen2.5-Coder-1.5B, hook layer 6 residuals.
3. Train TopK SAE via SA... |
eren23/coder-interp-tap | launchers/nla_coder/delta.py | py | 1,095 | 6d228cb2a708a493b3357dbd8f538005b7ee45db9c75df432d10043533244480 | """Phase-3 LoRA delta study launcher for nla_qwen2_5_coder_1_5b.
SCAFFOLD ONLY. Real implementation goes here once Phase 2 (SAE + AV) ships.
Intended pipeline:
1. LoRA-tune Qwen2.5-Coder-1.5B on a small fine-tune signal dataset
(synthetic preference pairs OR external preference data SCP'd in via
Crucible ... |
eren23/coder-interp-tap | launchers/nla_coder/smoke.py | py | 2,384 | 06936681abcd61de5a708ce7f5a4e4fe8702ae1807427ecd2052fa1d4a9f3c93 | """Phase-0 smoke launcher for nla_qwen2_5_coder_1_5b.
Same pattern as launchers/nla_pilot/smoke.py — validates pod plumbing,
logs a fake training curve to W&B, exits 0.
"""
from __future__ import annotations
import os
import time
def _env(name: str, default: str | None = None) -> str:
val = os.environ.get(name... |
eren23/coder-interp-tap | launchers/nla_pilot/pilot.py | py | 1,174 | 46469a7ac015a6831babeb9d38f582288b7c3cc1c9507d07b205921511a89d8f | """Phase-1 pilot launcher for nla_qwen3_5_2b_pilot.
SCAFFOLD ONLY. Real implementation goes here when Phase 0 smoke is green.
Intended pipeline:
1. residual_dump — forward 200K tokens through Qwen3.5-2B, hook layer 12,
cache (token_id, residual) to data/qwen3_5_2b_residuals_layer12.h5
2. ... |
eren23/coder-interp-tap | launchers/nla_pilot/smoke.py | py | 2,691 | c4080cdb07a0bb289ec04fd96878215f35a9de8c16ba1fee4d8c5ca177b41e97 | """Phase-0 smoke launcher for nla_qwen3_5_2b_pilot.
Validates pod plumbing:
- reads project env vars
- logs to W&B
- sleeps briefly
- exits 0
NO real training. Replace with the actual Phase-1 pilot loop once the
project YAML's `pilot` variant becomes the focus.
"""
from __future__ import annotations
import ... |
eren23/coder-interp-tap | launchers/feature_explainer/main.py | py | 18,420 | b8718a334d95f33e9f57234c1427ce0b37527cfa1ccaff3669a749edf7f228b6 | """Feature explainer — auto-interpret SAE features via an LLM.
For each SAE feature, collect its top-K firing contexts from a streaming
corpus, then ask an OpenAI-compatible LLM (default: DeepSeek V3 via
OpenRouter) to describe the concept the feature represents. Output goes
to a JSON file + a W&B Table.
Hot-paths:
... |
eren23/coder-interp-tap | launchers/lora_dpo/main.py | py | 4,169 | 6b22e3ac1e5857c3d857b5fa44e666abb90a072ebd25d483c04793b2c307a9de | """LoRA + DPO fine-tune launcher — STUB.
This launcher validates pod plumbing (HF cache layout, W&B init, model
download path) but does NOT yet run the actual TRL DPOTrainer. Real
implementation goes here when you start Track B in earnest.
Outline of the real pipeline (TODO):
1. Build a small synthetic preference d... |
eren23/coder-interp-tap | launchers/nla_real/pilot.py | py | 14,005 | 97e220cdc36de9c5cb7e3e170ce1d80e2785f03fe979d7a12d60a291a3bf3c0a | """Real NLA pilot launcher (transformers-only, no sglang).
End-to-end pipeline that produces actual natural-language descriptions of
Qwen2.5-7B-Instruct activation vectors via the kitft Activation Verbalizer
at layer 20.
The injection protocol (per kitft's nla_meta.yaml):
- prompt template wraps the ㈎ injection cha... |
eren23/coder-interp-tap | launchers/nla_real/smoke.py | py | 607 | 4c719cece1f5db09136262c70306ed8f48d6be2814acba37ff312f6a8e7e0702 | """Smoke variant — same pipeline as pilot, gated by env (NUM_PROMPTS=3,
POSITIONS_PER_PROMPT=1 in the project YAML's smoke variant block).
Crucible runs us as `python3 -u launchers/nla_real/smoke.py` so the
package import path "launchers.nla_real.pilot" is NOT on sys.path. Add
this script's parent dir to sys.path so `... |
eren23/coder-interp-tap | launchers/feature_diff_study/main.py | py | 25,441 | 4217c81619b0ef78fa44570c3b06c01e04e8624441034c439b750192e3d54325 | """Feature-diff delta study — end-to-end demo of the Track-C measurement pipeline.
Pipeline (all phases run in sequence on one pod):
1. Load base Qwen2.5-Coder-1.5B.
2. Build a small biased SFT dataset from a CommitPackFT slice.
3. LoRA-SFT for LORA_TRAIN_STEPS steps. Save adapter.
4. Forward HOLDOUT_TOKENS he... |
eren23/coder-interp-tap | launchers/feature_diff_study/summarize_delta.py | py | 9,242 | 042551f6a8ed468c3b206d44c580a6f184562f3b715ded38fdee2e81ab1d7034 | """LLM-explains-delta postprocess for feature_diff_study.
Compresses the top-K drift table into a 2-3 paragraph natural-language
summary of what the LoRA fine-tune actually did. Same OpenAI-compatible
endpoint pattern as feature_explainer (default DeepSeek V3 via OpenRouter).
Drop-in: imported by main.py at end of ru... |
eren23/coder-interp-tap | launchers/sae_train/main.py | py | 11,786 | 4d233b2277ff650d06d52792b0815246380464fa1e330f4311f4868cb2d76614 | """TopK Sparse Autoencoder trainer over a HuggingFace base model's
residual stream.
Pipeline:
1. Load base model (BASE_MODEL) in bf16 on cuda.
2. Stream documents from a HF dataset (DATASET_ID / DATASET_SUBSET) and
tokenize on the fly to SEQ_LEN.
3. Forward CAPTURE_BATCH_SIZE sequences at a time; capture re... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.