Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
tags: [interpretability, credit-assignment, steering-vectors, alignment]
|
| 4 |
+
---
|
| 5 |
+
|
| 6 |
+
# Credit-assignment artifacts
|
| 7 |
+
|
| 8 |
+
Weights/vectors, datasets, and per-model results for the **narrated credit-assignment** study.
|
| 9 |
+
Code, figures, and the interactive Craftax viz live on GitHub: `DavidDemitriAfrica/hackasack`.
|
| 10 |
+
Companion to [`davidafrica/functional-wellbeing`](https://huggingface.co/davidafrica/functional-wellbeing)
|
| 11 |
+
(the welfare axis used here).
|
| 12 |
+
|
| 13 |
+
## Contents
|
| 14 |
+
- `artifacts/credit_v1_*` — `v_credit.pt` credit direction per model (9 families, 1.7B–35B), plus
|
| 15 |
+
`analysis.json`, `steer.json`, `miscredit.json`, `repeat_choice.json`, `align.json`, `labels.parquet`.
|
| 16 |
+
- `artifacts/credit_v2*_*`, `maze_*` — confound (raw-rate-hard) + ecological-maze runs with `stats.json`.
|
| 17 |
+
- `artifacts/conflict_*`, `reasoning_*`, `behavioral_axis_*` — the structural-vs-behavioral / framing results
|
| 18 |
+
(incl. `v_struct.npy`, `v_behav.npy`).
|
| 19 |
+
- `artifacts/train_*`, `rlpay_*`, `syco_*`, `craftax_*` — SFT/RL training, sycophancy, and Craftax results.
|
| 20 |
+
- `datasets/*.parquet` — the generated stimulus sets (credit_v1/v2, maze, craftax).
|
| 21 |
+
|
| 22 |
+
Raw `activations.pt` were not retained (large; regenerable from the datasets via
|
| 23 |
+
`credit.extract_credit` in the GitHub repo).
|