runs/molab: logs
Browse files
runs/molab/logs/console.log
CHANGED
|
@@ -219,3 +219,11 @@ Consider using tensor.detach() first. (Triggered internally at /pytorch/torch/cs
|
|
| 219 |
18:19:00 | INFO | transcoder llama-13b step 2600 | 3.99M/4.00M tokens | loss 1.8383 fvu 0.316 dead 0.001 | 1575 tok/s
|
| 220 |
18:20:27 | INFO | transcoder llama-13b done: mean held-out FVU 0.355 over 5 layers (4001280 tokens)
|
| 221 |
18:25:10 | INFO | graph llama-13b/mod_arith: 2592 vertices (of 204800 features), 9915 edges | nodes 2.9s edges 222.6s
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 219 |
18:19:00 | INFO | transcoder llama-13b step 2600 | 3.99M/4.00M tokens | loss 1.8383 fvu 0.316 dead 0.001 | 1575 tok/s
|
| 220 |
18:20:27 | INFO | transcoder llama-13b done: mean held-out FVU 0.355 over 5 layers (4001280 tokens)
|
| 221 |
18:25:10 | INFO | graph llama-13b/mod_arith: 2592 vertices (of 204800 features), 9915 edges | nodes 2.9s edges 222.6s
|
| 222 |
+
18:34:01 | INFO | circuit llama-13b/mod_arith/hierarchical: 849 nodes, 6489 edges, phi=0.945 (full acc 1.000), 686s
|
| 223 |
+
18:34:57 | INFO | circuit llama-13b/mod_arith/attribution_patching: 1020 nodes, 6015 edges, phi=0.926 (full acc 1.000), 240s
|
| 224 |
+
18:35:52 | INFO | circuit llama-13b/mod_arith/sparse_probing: 1020 nodes, 0 edges, phi=0.926 (full acc 1.000), 18s
|
| 225 |
+
18:35:54 | INFO | ACDC llama-13b/mod_arith: the 12-feature/layer pool itself scores -0.073 < theta; pruning cannot recover it
|
| 226 |
+
18:37:13 | INFO | ACDC llama-13b/mod_arith: 588 -> 0 edges, 4 vertices (ok, 81s)
|
| 227 |
+
18:37:54 | INFO | circuit llama-13b/mod_arith/acdc: 4 nodes, 0 edges, phi=0.000 (full acc 1.000), 84s
|
| 228 |
+
18:37:57 | INFO | llama-13b/parity: model accuracy 0.514 is at chance (0.500; needs 0.550) - no circuit search
|
| 229 |
+
18:46:40 | INFO | graph llama-13b/sorting: 2592 vertices (of 204800 features), 12865 edges | nodes 3.7s edges 486.5s
|
runs/molab/logs/pipeline.log
CHANGED
|
@@ -192,3 +192,11 @@ Make sure to have access to it at https://huggingface.co/meta-llama/Llama-2-13b-
|
|
| 192 |
18:19:00 | INFO | transcoder llama-13b step 2600 | 3.99M/4.00M tokens | loss 1.8383 fvu 0.316 dead 0.001 | 1575 tok/s
|
| 193 |
18:20:27 | INFO | transcoder llama-13b done: mean held-out FVU 0.355 over 5 layers (4001280 tokens)
|
| 194 |
18:25:10 | INFO | graph llama-13b/mod_arith: 2592 vertices (of 204800 features), 9915 edges | nodes 2.9s edges 222.6s
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 192 |
18:19:00 | INFO | transcoder llama-13b step 2600 | 3.99M/4.00M tokens | loss 1.8383 fvu 0.316 dead 0.001 | 1575 tok/s
|
| 193 |
18:20:27 | INFO | transcoder llama-13b done: mean held-out FVU 0.355 over 5 layers (4001280 tokens)
|
| 194 |
18:25:10 | INFO | graph llama-13b/mod_arith: 2592 vertices (of 204800 features), 9915 edges | nodes 2.9s edges 222.6s
|
| 195 |
+
18:34:01 | INFO | circuit llama-13b/mod_arith/hierarchical: 849 nodes, 6489 edges, phi=0.945 (full acc 1.000), 686s
|
| 196 |
+
18:34:57 | INFO | circuit llama-13b/mod_arith/attribution_patching: 1020 nodes, 6015 edges, phi=0.926 (full acc 1.000), 240s
|
| 197 |
+
18:35:52 | INFO | circuit llama-13b/mod_arith/sparse_probing: 1020 nodes, 0 edges, phi=0.926 (full acc 1.000), 18s
|
| 198 |
+
18:35:54 | INFO | ACDC llama-13b/mod_arith: the 12-feature/layer pool itself scores -0.073 < theta; pruning cannot recover it
|
| 199 |
+
18:37:13 | INFO | ACDC llama-13b/mod_arith: 588 -> 0 edges, 4 vertices (ok, 81s)
|
| 200 |
+
18:37:54 | INFO | circuit llama-13b/mod_arith/acdc: 4 nodes, 0 edges, phi=0.000 (full acc 1.000), 84s
|
| 201 |
+
18:37:57 | INFO | llama-13b/parity: model accuracy 0.514 is at chance (0.500; needs 0.550) - no circuit search
|
| 202 |
+
18:46:40 | INFO | graph llama-13b/sorting: 2592 vertices (of 204800 features), 12865 edges | nodes 3.7s edges 486.5s
|