AlexWortega commited on
Commit
f1a5602
·
verified ·
1 Parent(s): 378a074

Retitle: Weight-Space Geometry of Offline Reasoning Training

Browse files
Files changed (3) hide show
  1. README.md +4 -4
  2. data/meta.json +2 -2
  3. index.html +1 -1
README.md CHANGED
@@ -1,5 +1,5 @@
1
  ---
2
- title: Same Data, Different Losses, Same Circuits?
3
  emoji: 🧭
4
  colorFrom: indigo
5
  colorTo: pink
@@ -10,10 +10,10 @@ license: mit
10
  short_description: Interactive weight-space geometry of six reasoning losses
11
  ---
12
 
13
- # Same Data, Different Losses, Same Circuits?
14
 
15
- An **interactive blog** companion to the ICML 2026 MechInterp Workshop paper of the same
16
- name. Six offline reasoning losses (SFT, RFT, DFT, RIFT, Offline GRPO, DPO) are trained on
17
  *identical* math rollouts from one base model (Qwen3-4B, attention-only LoRA). The page lets
18
  you turn the same knobs the authors did — pick method pairs, scrub across all 36 transformer
19
  layers, and watch the geometry respond.
 
1
  ---
2
+ title: Weight-Space Geometry of Offline Reasoning Training
3
  emoji: 🧭
4
  colorFrom: indigo
5
  colorTo: pink
 
10
  short_description: Interactive weight-space geometry of six reasoning losses
11
  ---
12
 
13
+ # Weight-Space Geometry of Offline Reasoning Training
14
 
15
+ An **interactive blog** companion to the ICML 2026 MechInterp Workshop paper
16
+ (*"Same Data, Different Losses, Same Circuits?"*). Six offline reasoning losses (SFT, RFT, DFT, RIFT, Offline GRPO, DPO) are trained on
17
  *identical* math rollouts from one base model (Qwen3-4B, attention-only LoRA). The page lets
18
  you turn the same knobs the authors did — pick method pairs, scrub across all 36 transformer
19
  layers, and watch the geometry respond.
data/meta.json CHANGED
@@ -1,6 +1,6 @@
1
  {
2
- "title": "Same Data, Different Losses, Same Circuits?",
3
- "subtitle": "Weight-space geometry of six offline reasoning losses trained on identical math rollouts from one base model.",
4
  "venue": "ICML 2026 · Mechanistic Interpretability Workshop",
5
  "repo": "https://github.com/AlexWortega/capabilityvectors",
6
  "methodColors": {
 
1
  {
2
+ "title": "Weight-Space Geometry of Offline Reasoning Training",
3
+ "subtitle": "Same data, different losses, same circuits? Six offline reasoning losses, one base model, identical math rollouts — do they converge in weight space?",
4
  "venue": "ICML 2026 · Mechanistic Interpretability Workshop",
5
  "repo": "https://github.com/AlexWortega/capabilityvectors",
6
  "methodColors": {
index.html CHANGED
@@ -3,7 +3,7 @@
3
  <head>
4
  <meta charset="utf-8" />
5
  <meta name="viewport" content="width=device-width, initial-scale=1" />
6
- <title>Same Data, Different Losses, Same Circuits?</title>
7
  <meta name="description" content="An interactive look at the weight-space geometry of six offline reasoning losses trained on identical data." />
8
  <link rel="preconnect" href="https://fonts.googleapis.com" />
9
  <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin />
 
3
  <head>
4
  <meta charset="utf-8" />
5
  <meta name="viewport" content="width=device-width, initial-scale=1" />
6
+ <title>Weight-Space Geometry of Offline Reasoning Training</title>
7
  <meta name="description" content="An interactive look at the weight-space geometry of six offline reasoning losses trained on identical data." />
8
  <link rel="preconnect" href="https://fonts.googleapis.com" />
9
  <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin />