Add the formality steering grid on GSM8K
Browse files- README.md +16 -0
- formality-gsm8k-grid/README.md +40 -0
- formality-gsm8k-grid/grid.json +0 -0
README.md
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
pretty_name: Steerability optimization data
|
| 4 |
+
viewer: false
|
| 5 |
+
tags:
|
| 6 |
+
- steerability
|
| 7 |
+
- activation-steering
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
# Steerability optimization data
|
| 11 |
+
|
| 12 |
+
Recorded data of the optimization experiments of the [Steerability toolkit](https://github.com/generative-computing/steerability). Each folder contains one recorded experiment, together with a `README.md` that describes its contents and how it was recorded. The notebooks of the toolkit download these files at a pinned commit of this repository.
|
| 13 |
+
|
| 14 |
+
| Folder | Contents | Read by |
|
| 15 |
+
|---|---|---|
|
| 16 |
+
| [`formality-gsm8k-grid/`](./formality-gsm8k-grid/README.md) | Per-item scores of `Qwen/Qwen2.5-0.5B-Instruct` under one CAA formality vector at 10 layers and 10 multipliers, and of the unsteered model, on GSM8K and a formality objective | `examples/notebooks/optimization/optimization.ipynb` |
|
formality-gsm8k-grid/README.md
ADDED
|
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Formality steering grid
|
| 2 |
+
|
| 3 |
+
This folder contains the per-item scores of `Qwen/Qwen2.5-0.5B-Instruct` under one CAA formality vector at 10 layers and 10 multipliers, and of the unsteered model, on two benchmarks. Its `grid.json` is the input of the optimization notebook of the [Steerability toolkit](https://github.com/generative-computing/steerability).
|
| 4 |
+
|
| 5 |
+
## Contents
|
| 6 |
+
|
| 7 |
+
`grid.json` is one JSON object with two keys:
|
| 8 |
+
|
| 9 |
+
- `settings`: the 100 settings as `{"caa.layer_id": <int>, "caa.multiplier": <float>}`, in the order of the per-setting measurements below.
|
| 10 |
+
- `benchmarks`: one entry for GSM8K (`"capability"`) and one for the formality objective (`"behavior"`). Each entry has the item ids (`items`, 3,792 GSM8K ids and 1,500 prompt ids), two measurements of the unsteered model in different batch orders (`baseline`), and one measurement per setting (`settings`).
|
| 11 |
+
|
| 12 |
+
A measurement has one entry per item, in the order of `items`. Its lists are `values` (the GSM8K correctness in {0, 1}, or the formality score in [0, 1]) and `tokens` (the number of completion tokens). The measurements of the formality objective also have `degenerate`, which flags each degenerate completion. The file contains no prompts and no completions.
|
| 13 |
+
|
| 14 |
+
## How it was recorded
|
| 15 |
+
|
| 16 |
+
- Model and protocol: `Qwen/Qwen2.5-0.5B-Instruct`, float32, greedy decoding (temperature 0, seed 0), batches of 64, Hugging Face backend, steerability 0.6.0 with torch 2.13.0 and transformers 5.19.0 (protocol id `af8f630d8b761825`).
|
| 17 |
+
- Capability benchmark: GSM8K, train and test splits (8,792 questions), zero-shot, with at most 384 completion tokens. The grid uses the 3,792 search items that remain after 5,000 items are sealed (seed 0).
|
| 18 |
+
- Objective: formality of the replies to 1,500 of 2,100 short messaging prompts (600 prompts are sealed, seed 0), with at most 96 completion tokens, scored by [`s-nlp/roberta-base-formality-ranker`](https://huggingface.co/s-nlp/roberta-base-formality-ranker).
|
| 19 |
+
- Steering vector: the mean difference between the activations of eight formal messages and their casual versions.
|
| 20 |
+
- Settings: layers 6, 8, 10, 12, 14, 16, 18, 20, 22, 23 and multipliers 0.5, 1, 1.5, 2, 3, 4, 5, 6, 8, 10. The unsteered model was measured twice.
|
| 21 |
+
- Commands: the settings were divided into eight shards, and shard 7 also measured the unsteered model (the other shards passed `--skip-baseline`). GSM8K and the formality objective were measured in separate passes over the shards. The GSM8K pass ran an earlier revision of `truth_grid.py` that measured GSM8K only, under the same protocol.
|
| 22 |
+
|
| 23 |
+
python examples/notebooks/optimization/scripts/truth_grid.py collect runs/truth_grid/shard_<i> --shard <i> --shards 8 --benchmarks gsm8k --skip-baseline
|
| 24 |
+
python examples/notebooks/optimization/scripts/truth_grid.py collect runs/truth_grid/shard_<i> --shard <i> --shards 8 --benchmarks formality --skip-baseline
|
| 25 |
+
python examples/notebooks/optimization/scripts/replay.py export runs/truth_grid/shard_{0..7} --output grid.json
|
| 26 |
+
|
| 27 |
+
- Cost: about 21 GPU-hours on NVIDIA A100 GPUs (about 19 for GSM8K and 2 for the formality objective).
|
| 28 |
+
|
| 29 |
+
## Use
|
| 30 |
+
|
| 31 |
+
from huggingface_hub import hf_hub_download
|
| 32 |
+
|
| 33 |
+
path = hf_hub_download(
|
| 34 |
+
repo_id="Steerability/optimization",
|
| 35 |
+
filename="formality-gsm8k-grid/grid.json",
|
| 36 |
+
repo_type="dataset",
|
| 37 |
+
revision="<a commit of this repository>",
|
| 38 |
+
)
|
| 39 |
+
|
| 40 |
+
Pin a commit, as the optimization notebook does. `replay.load_export(path)` in the toolkit's `examples/notebooks/optimization/scripts/replay.py` reads the file.
|
formality-gsm8k-grid/grid.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|