Pacific-i64 commited on
Commit
b8a4dbb
·
verified ·
1 Parent(s): 59ec633

Update live comparison for TR-Hash 500M

Browse files
Files changed (2) hide show
  1. README.md +5 -3
  2. index.html +6 -3
README.md CHANGED
@@ -19,13 +19,15 @@ The Space provides:
19
  - an architecture explorer for the shared SwiGLU and fixed top-2 residual path;
20
  - checkpoint-derived tokenizer IDs and expert assignments for selected tokens;
21
  - the reported matched-run measurements and their stated limitations;
22
- - live side-by-side SSE generation from `Pacific-i64/TR-MOE-306` and `Pacific-i64/Dense-306`.
23
 
24
- The interactive interface is a presentation and inspection companion. It is not additional experimental evidence.
 
 
25
 
26
  ## Sources
27
 
28
  - [Paper and reproducibility artifacts](https://github.com/Complexity-ML/tmlr-paper-pool)
29
- - [TR-MOE-306](https://huggingface.co/Pacific-i64/TR-MOE-306)
30
  - [Dense-306](https://huggingface.co/Pacific-i64/Dense-306)
31
  - [vllm-i64](https://github.com/Complexity-ML/vllm-i64)
 
19
  - an architecture explorer for the shared SwiGLU and fixed top-2 residual path;
20
  - checkpoint-derived tokenizer IDs and expert assignments for selected tokens;
21
  - the reported matched-run measurements and their stated limitations;
22
+ - live side-by-side SSE generation from `Pacific-i64/TR-HASH-MOE-500M-HF` and `Pacific-i64/Dense-306`.
23
 
24
+ The historical measurements remain the matched 306.5M study. The live panel now
25
+ compares the current 492.1M TR-Hash release with Dense-306, so its outputs are a
26
+ non-iso-parameter qualitative inspection and not additional experimental evidence.
27
 
28
  ## Sources
29
 
30
  - [Paper and reproducibility artifacts](https://github.com/Complexity-ML/tmlr-paper-pool)
31
+ - [TR-HASH-MOE-500M-HF](https://huggingface.co/Pacific-i64/TR-HASH-MOE-500M-HF)
32
  - [Dense-306](https://huggingface.co/Pacific-i64/Dense-306)
33
  - [vllm-i64](https://github.com/Complexity-ML/vllm-i64)
index.html CHANGED
@@ -557,13 +557,13 @@
557
  </div>
558
  <div class="card">
559
  <div class="tr-run-head">
560
- <strong>TR-MOE-306</strong>
561
  <span class="viz-badge" data-routed-state>ready</span>
562
  </div>
563
  <p class="tr-generation" data-routed-output aria-live="polite">Token-routed output will stream here token by token.</p>
564
  </div>
565
  </div>
566
- <div class="tr-status text-small" data-status>Public CPU inference · SSE streaming · qualitative comparison only</div>
567
  </section>
568
  </div>
569
 
@@ -747,7 +747,7 @@
747
  denseState.textContent = "connecting";
748
  routedState.textContent = "connecting";
749
  runButton.textContent = "Stop generation";
750
- status.textContent = "Connecting to both public 306.5M CPU endpoints…";
751
  status.classList.add("is-live");
752
 
753
  try {
@@ -758,6 +758,9 @@
758
  prompt,
759
  max_tokens: 128,
760
  temperature: 0.7,
 
 
 
761
  stream: true
762
  }),
763
  signal: controller.signal
 
557
  </div>
558
  <div class="card">
559
  <div class="tr-run-head">
560
+ <strong>TR-HASH-MOE-500M</strong>
561
  <span class="viz-badge" data-routed-state>ready</span>
562
  </div>
563
  <p class="tr-generation" data-routed-output aria-live="polite">Token-routed output will stream here token by token.</p>
564
  </div>
565
  </div>
566
+ <div class="tr-status text-small" data-status>TR-Hash 492.1M vs Dense 306.5M · non-iso-parameter qualitative comparison</div>
567
  </section>
568
  </div>
569
 
 
747
  denseState.textContent = "connecting";
748
  routedState.textContent = "connecting";
749
  runButton.textContent = "Stop generation";
750
+ status.textContent = "Connecting to TR-Hash 492.1M and Dense 306.5M CPU endpoints…";
751
  status.classList.add("is-live");
752
 
753
  try {
 
758
  prompt,
759
  max_tokens: 128,
760
  temperature: 0.7,
761
+ top_k: 40,
762
+ top_p: 0.9,
763
+ repetition_penalty: 1.1,
764
  stream: true
765
  }),
766
  signal: controller.signal