AlekseyCalvin commited on
Commit
dcafe31
·
verified ·
1 Parent(s): 3ef05f8

Rename NLA tab to NLA · SHOW & TELL

Browse files
README.md CHANGED
@@ -203,7 +203,7 @@ datasets:
203
  - **Evaluation & Feature Dashboards**: Explained variance, CE recovered and KL score from spliced forward passes; activation densities, top contexts, logit effects, auto-interpretation, and a 3-D feature atlas.
204
  - **Concept Manifolds & Steering**: Fit a concept manifold (spline through PCA centroids) and steer along it, or steer, clamp and ablate single SAE features in the chat.
205
 
206
- - **◉ NLA · TALKING CURE**:
207
  - **Natural Language Autoencoders**: A verbalizer turns an activation into text and a reconstructor maps the text back to an activation. Each explanation is scored by reconstruction (cosine, FVE against a streamed reference corpus).
208
  - **Vector Modes & Position Scans**: Read single activations, differences, analogies, or model diffs against a fine-tune (with ActShift), and scan many positions in one batch.
209
  - **Rewrite & Steer**: Edit an explanation; the change in its reconstruction becomes a steering direction for the chat.
 
203
  - **Evaluation & Feature Dashboards**: Explained variance, CE recovered and KL score from spliced forward passes; activation densities, top contexts, logit effects, auto-interpretation, and a 3-D feature atlas.
204
  - **Concept Manifolds & Steering**: Fit a concept manifold (spline through PCA centroids) and steer along it, or steer, clamp and ablate single SAE features in the chat.
205
 
206
+ - **◉ NLA · SHOW & TELL**:
207
  - **Natural Language Autoencoders**: A verbalizer turns an activation into text and a reconstructor maps the text back to an activation. Each explanation is scored by reconstruction (cosine, FVE against a streamed reference corpus).
208
  - **Vector Modes & Position Scans**: Read single activations, differences, analogies, or model diffs against a fine-tune (with ActShift), and scan many positions in one batch.
209
  - **Rewrite & Steer**: Edit an explanation; the change in its reconstruction becomes a steering direction for the chat.
bistro/interp/__init__.py CHANGED
@@ -1,4 +1,4 @@
1
- """Interpretability, steering & autoencoder instruments (SAE · COANALYSIS, NLA · TALKING CURE, steering panel).
2
 
3
  Everything here works on whatever model the Bistro engine loads (any HF architecture, custom remote code,
4
  synthesized universal models), through forward hooks on the model's own residual blocks."""
 
1
+ """Interpretability, steering & autoencoder instruments (SAE · COANALYSIS, NLA · SHOW & TELL, steering panel).
2
 
3
  Everything here works on whatever model the Bistro engine loads (any HF architecture, custom remote code,
4
  synthesized universal models), through forward hooks on the model's own residual blocks."""
bistro/ui/glossary.py CHANGED
@@ -252,7 +252,7 @@ HINTS.update({
252
  "SAE Bundle (.zip)": "Zipped SAELens folder: `cfg.json`, `sae_weights.safetensors`, `sparsity.safetensors`, `bistro_sae.json`, `README.md`.",
253
  })
254
 
255
- # ── NLA · TALKING CURE ──
256
  HINTS.update({
257
  "autoencoder & subject model": "Natural language autoencoder (Fraser-Taliente et al. 2026): a verbalizer (AV) writes an explanation $z$ of an activation $h$, a reconstructor (AR) maps $z$ back to $\\hat h$; both are fine-tuned copies of the subject model, scored by how well $\\hat h$ recovers $h$.",
258
  "NLA Preset": "Hub NLAs in the reference layout (kitft/natural_language_autoencoders: `nla_meta.yaml` sidecar, safetensors) small enough for a CPU Space.",
@@ -313,7 +313,7 @@ HINTS.update({
313
  "Steering Vector (.safetensors)": "The last contrastive vector $v$ with its layer, site, prompts and model in the safetensors metadata.",
314
  })
315
 
316
- # ── NLA · TALKING CURE: activation oracle ──
317
  HINTS.update({
318
  "Activation Oracle (Karvonen et al. 2025)": "A LoRA on the model itself, trained to answer questions about its own activations (Karvonen et al. 2025, arXiv 2512.15674). Activations $v_1 \\dots v_N$ from layer $L$ are added at $N$ placeholder tokens of the question prompt, after layer 1: $h_i \\leftarrow h_i + \\|h_i\\|\\, v_i / \\|v_i\\|$.",
319
  "oracle": "LoRA and base model of an oracle checkpoint; `ao_config.json` gives its trained layers (25 / 50 / 75 % depth), the placeholder token and the injection layer. One copy of the base holds the unmerged LoRA: adapter off = subject model, adapter on = oracle.",
 
252
  "SAE Bundle (.zip)": "Zipped SAELens folder: `cfg.json`, `sae_weights.safetensors`, `sparsity.safetensors`, `bistro_sae.json`, `README.md`.",
253
  })
254
 
255
+ # ── NLA · SHOW & TELL ──
256
  HINTS.update({
257
  "autoencoder & subject model": "Natural language autoencoder (Fraser-Taliente et al. 2026): a verbalizer (AV) writes an explanation $z$ of an activation $h$, a reconstructor (AR) maps $z$ back to $\\hat h$; both are fine-tuned copies of the subject model, scored by how well $\\hat h$ recovers $h$.",
258
  "NLA Preset": "Hub NLAs in the reference layout (kitft/natural_language_autoencoders: `nla_meta.yaml` sidecar, safetensors) small enough for a CPU Space.",
 
313
  "Steering Vector (.safetensors)": "The last contrastive vector $v$ with its layer, site, prompts and model in the safetensors metadata.",
314
  })
315
 
316
+ # ── NLA · SHOW & TELL: activation oracle ──
317
  HINTS.update({
318
  "Activation Oracle (Karvonen et al. 2025)": "A LoRA on the model itself, trained to answer questions about its own activations (Karvonen et al. 2025, arXiv 2512.15674). Activations $v_1 \\dots v_N$ from layer $L$ are added at $N$ placeholder tokens of the question prompt, after layer 1: $h_i \\leftarrow h_i + \\|h_i\\|\\, v_i / \\|v_i\\|$.",
319
  "oracle": "LoRA and base model of an oracle checkpoint; `ao_config.json` gives its trained layers (25 / 50 / 75 % depth), the placeholder token and the injection layer. One copy of the base holds the unmerged LoRA: adapter off = subject model, adapter on = oracle.",
bistro/ui/nla_tab.py CHANGED
@@ -1,4 +1,4 @@
1
- """Tab 4: ◉ NLA · TALKING CURE — natural language autoencoders on the live model: read activations into words
2
  (single vectors, differences, analogies, model diffs), score every explanation by reconstruction, scan positions,
3
  and steer the chat by rewriting an explanation."""
4
  import html as _html
@@ -152,7 +152,7 @@ def _status(extra: str = "") -> str:
152
  def _overview_html() -> str:
153
  nla = N.NLA_REG.nla
154
  if nla is None:
155
- return empty_html("NLA · TALKING CURE", "Load a verbalizer (AV) and reconstructor (AR); explanations render here.")
156
  m = nla.meta
157
  n_av = sum(p.numel() for p in nla.av.parameters())
158
  dt = str(next(nla.av.parameters()).dtype).replace("torch.", "")
@@ -333,7 +333,7 @@ def build_nla_tab(telem_html_comp: gr.HTML, bw: SimpleNamespace) -> SimpleNamesp
333
  first = next(iter(N.NLA_PRESETS))
334
  p0 = N.NLA_PRESETS[first]
335
 
336
- with gr.Tab("◉ NLA · TALKING CURE"):
337
  with split():
338
  with ctl():
339
  sub("autoencoder & subject model")
 
1
+ """Tab 4: ◉ NLA · SHOW & TELL — natural language autoencoders on the live model: read activations into words
2
  (single vectors, differences, analogies, model diffs), score every explanation by reconstruction, scan positions,
3
  and steer the chat by rewriting an explanation."""
4
  import html as _html
 
152
  def _overview_html() -> str:
153
  nla = N.NLA_REG.nla
154
  if nla is None:
155
+ return empty_html("NLA · SHOW & TELL", "Load a verbalizer (AV) and reconstructor (AR); explanations render here.")
156
  m = nla.meta
157
  n_av = sum(p.numel() for p in nla.av.parameters())
158
  dt = str(next(nla.av.parameters()).dtype).replace("torch.", "")
 
333
  first = next(iter(N.NLA_PRESETS))
334
  p0 = N.NLA_PRESETS[first]
335
 
336
+ with gr.Tab("◉ NLA · SHOW & TELL"):
337
  with split():
338
  with ctl():
339
  sub("autoencoder & subject model")
bistro/ui/oracle_panel.py CHANGED
@@ -1,4 +1,4 @@
1
- """Activation Oracle panel at the foot of ◉ NLA · TALKING CURE (Karvonen et al. 2025): ask a LoRA-tuned copy of a
2
  model free-form questions about its own activations, from one token, a window, every token in turn, or the whole text;
3
  activations from the oracle's base model, the BISTRO model (e.g. a fine-tune), or their difference."""
4
  import html as _html
 
1
+ """Activation Oracle panel at the foot of ◉ NLA · SHOW & TELL (Karvonen et al. 2025): ask a LoRA-tuned copy of a
2
  model free-form questions about its own activations, from one token, a window, every token in turn, or the whole text;
3
  activations from the oracle's base model, the BISTRO model (e.g. a fine-tune), or their difference."""
4
  import html as _html