tsce_demo

KO | EN

gitlite — search

#react #ai #typescript #go #deep-learning #docker #macos #java #windows #awesome #awesome-list #vue #security #mcp

(★ 32)

This repo is for the demonstration of TSCE principles.

tsce_demo 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# TSCE Paper Companion Repository

> **New (2026-07-09): [`jlens_bridge/`](jlens_bridge/README.md)** — connecting TSCE to Anthropic's
> [global-workspace / J-lens finding](https://transformer-circuits.pub/2026/workspace/index.html):
> the paper's published (blinded) HDA anchors load the hidden target concept's J-lens directions
> pre-decode on Gemma (**anchor − matched-random = +0.0150, CI [+0.0121, +0.0181], n=40**), with the
> full condition ordering matching the paper's behavioral probe margins and the effect localizing
> at the layers Figure 6 flagged independently. Code, results, and reproduction steps inside.

This repository is a self-contained companion artifact for the TSCE paper. It includes the paper source and PDF, the paper-facing test sets, curated result artifacts that back the paper tables and figures, and the installable TSCE wrapper package. It intentionally excludes caches, virtual environments, generated bytecode, dashboards, model checkpoints, broad exploratory runs, and unrelated training outputs.

## Layout

- `paper/`: final `think_before_you_speak_v2` PDF/TEX, mechanism addendum, diffusion framework, support notes, referenced figures, and artifact-local figure-generation code.
- `test_sets/`: HDA task-suite eval/smoke/baseline-failure sets, GSM/GSM-Hard references, synthetic format-control eval/smoke sets, manifests, and HDA candidate files used by the format sweep.
- `results/`: curated paper-supporting result folders plus `docs/tsce_results_inventory.csv` and `docs/tsce_results_inventory_summary.json` as full-corpus provenance.
- `package/`: minimal installable TSCE wrapper package, examples, and wrapper tests.
- `scripts/`: artifact-local validation and figure-regeneration entry points.
- `MANIFEST.md`: source-to-destination provenance for every copied or generated artifact except the checksum file.
- `checksums.sha256`: SHA-256 hashes for the exported files.
- `SECURITY_REVIEW.md`: credential/key scan summary for publication.

## Quick Validation

From the repository root:

```bash
python scripts/validate_artifact.py
shasum -a 256 -c checksums.sha256
```

The validation script checks for banned artifacts, required paper/package/result files, checksum integrity, paper figure references, artifact-local executable paths, absence of local absolute paths, strict credential signatures, and package importability.

## Wrapper Install

```bash
cd package
python -m pip install -e .
python -c "from tsce import TSCEChat, TSCEClient, TSCEReply"
```

Wrapper tests from `package/`:

```bash
python -m pytest tests/test_tsce_wrapper.py tests/test_tsce_chat_smoke.py
```

If `pytest`, `numpy`, or `matplotlib` are missing, install optional local reproduction dependencies from the repo root:

```bash
python -m pip install -r requirements-reproduce.txt
```

## Paper Build

The checked-in PDF is `paper/think_before_you_speak_v2.pdf`. The TEX source uses figure paths relative to `paper/`, so a LaTeX engine can build it from that directory. For example, with Tectonic installed:

```bash
cd paper
tectonic --outdir /tmp/tsce-paper-build think_before_you_speak_v2.tex
```

## Regenerate Figures

The figure-generation script has been relocated to use this repository layout only. It reads the curated results under `results/` and writes to `paper/figures/tsce/`:

```bash
python scripts/build_figures.py
```

## Result Mapping

- Fixed-policy Gemma neutral validation: `results/tsce_agent_demo/results/gemma_fixed_latent_bias_compact_heldout_n100_seed260426/`.
- Format-robustness sweep on 69 baseline failures: `results/tsce_agent_demo/results/hda_anchor_format_full_sweep_seed260501_u69_k16_combined/`.
- Controlled greenhouse proxy-family and token-causality run: `results/tsce_agent_demo/results/story_subject_greenhouse_v1/`.
- Naturalistic forgiveness-theme token-causality run: `results/tsce_agent_demo/results/story_theme_forgiveness_v1/`.
- Activation-space diagnostic used by figure 6: `results/tsce_agent_demo/results/gemma_latent_zone_v1_20260424_progress2/`.
- Cross-run black-box figure inputs: curated `archive/2025-05-30_*`, `results-100-*`, `results-300-*`, `tsce_diffusion_2025-12-03_*`, `tsce_eval`, and `tsce_eval_deterministic` folders under `results/`.

## Integrity

`checksums.sha256` covers every tracked artifact file except itself. Re-run the checksum command after any intentional edit, then regenerate the checksum file before publishing.

tsce_demo

tsce_demo 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

tsce_demo

tsce_demo 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README