Mandeep Sidhu commited on
Commit ·
cf52b0e
1
Parent(s): dcae82e
Add WikiText-103 five-seed streaming validation
Browse files- docs/plan.md +55 -13
- docs/wikitext103_streaming_report.md +198 -0
- runs/wikitext103_l12_streaming_validation_5seed/locked_stream/20260531-093525/RESULT_SUMMARY.md +122 -0
- runs/wikitext103_l12_streaming_validation_5seed/locked_stream/20260531-093525/config.json +199 -0
- runs/wikitext103_l12_streaming_validation_5seed/locked_stream/20260531-093525/metrics.jsonl +0 -0
- runs/wikitext103_l12_streaming_validation_5seed/locked_stream/20260531-093525/summary.csv +71 -0
- runs/wikitext103_l12_streaming_validation_5seed/locked_stream/20260531-093525/summary.json +1542 -0
- runs/wikitext103_l12_streaming_validation_5seed/locked_stream/20260531-093525/trace.jsonl +0 -0
- runs/wikitext103_streaming_report/l12_validation_5seed/condition_summary.csv +15 -0
- runs/wikitext103_streaming_report/l12_validation_5seed/paired_final_deltas.csv +71 -0
- runs/wikitext103_streaming_report/l12_validation_5seed/stage_summary.csv +71 -0
docs/plan.md
CHANGED
|
@@ -285,7 +285,7 @@ Use this order for every regime.
|
|
| 285 |
| TinyStories static/coefficient regime | active | main coefficient evidence |
|
| 286 |
| TinyStories streaming regime | 5-seed validation complete | current main streaming evidence; interaction decay beats best static in 5/5 paired final-loss comparisons |
|
| 287 |
| OpenWebText10K streaming regime | 5-seed clean validation complete | OpenWebText10K interaction decay beats best static in 5/5 paired final-loss comparisons |
|
| 288 |
-
| WikiText-103 streaming regime |
|
| 289 |
|
| 290 |
## Current Formula Status
|
| 291 |
|
|
@@ -334,6 +334,7 @@ structure transfers, while coefficients may be regime-specific.
|
|
| 334 |
| TinyStories held-out model | supports pressure dependence on model size |
|
| 335 |
| TinyStories streaming, 5 seeds | interaction has best mean final loss; interaction beats best static in 5/5 paired final-loss comparisons |
|
| 336 |
| OpenWebText10K streaming, 5 seeds | interaction decay has best mean final loss; top decay schedules beat best static in 5/5 paired comparisons |
|
|
|
|
| 337 |
| cross-regime raw coefficient transfer | weaker than within-regime fit; supports regime-specific coefficients rather than universal numeric coefficients |
|
| 338 |
|
| 339 |
Latest TinyStories 5-seed streaming final-loss table:
|
|
@@ -401,6 +402,38 @@ The best static baseline in the clean OpenWebText10K run is static dropout
|
|
| 401 |
`0.0473` and wins every paired seed comparison. This promotes OpenWebText10K
|
| 402 |
from exploratory support to a second multi-seed streaming validation regime.
|
| 403 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 404 |
## Completed Static Backtest Gate
|
| 405 |
|
| 406 |
The first offline coefficient backtest is complete. It is retained as supporting
|
|
@@ -423,11 +456,15 @@ streaming multi-seed reports for each regime.
|
|
| 423 |
|
| 424 |
## Immediate Next Action
|
| 425 |
|
| 426 |
-
Reconcile the TinyStories
|
| 427 |
-
into the paper outline. The
|
| 428 |
-
|
| 429 |
-
|
| 430 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 431 |
|
| 432 |
## Next Training After Current Gate
|
| 433 |
|
|
@@ -438,8 +475,10 @@ limiting issue, use a third held-out regime for the next validation step:
|
|
| 438 |
```text
|
| 439 |
completed: TinyStories 5-seed streaming report
|
| 440 |
completed: OpenWebText10K 5-seed clean streaming report
|
| 441 |
-
|
| 442 |
-
|
|
|
|
|
|
|
| 443 |
```
|
| 444 |
|
| 445 |
Evaluate with paired seed comparisons:
|
|
@@ -452,17 +491,20 @@ decay minus best-static delta per seed
|
|
| 452 |
rank consistency across seeds
|
| 453 |
```
|
| 454 |
|
| 455 |
-
Because OpenWebText10K
|
| 456 |
-
streaming claim to "supported in
|
| 457 |
-
|
| 458 |
-
|
| 459 |
-
|
|
|
|
| 460 |
|
| 461 |
Latest streaming report:
|
| 462 |
|
| 463 |
```text
|
| 464 |
docs/tinystories_streaming_report.md
|
| 465 |
docs/openwebtext10k_streaming_report.md
|
|
|
|
| 466 |
runs/streaming_tinystories_multiseed_validation_l12/combined_5seed_summary/
|
| 467 |
runs/openwebtext10k_streaming_report/l16_updated_formula_clean_5seed/
|
|
|
|
| 468 |
```
|
|
|
|
| 285 |
| TinyStories static/coefficient regime | active | main coefficient evidence |
|
| 286 |
| TinyStories streaming regime | 5-seed validation complete | current main streaming evidence; interaction decay beats best static in 5/5 paired final-loss comparisons |
|
| 287 |
| OpenWebText10K streaming regime | 5-seed clean validation complete | OpenWebText10K interaction decay beats best static in 5/5 paired final-loss comparisons |
|
| 288 |
+
| WikiText-103 streaming regime | 5-seed validation complete | formula-derived L12 decay beats best static in 5/5 paired final-loss comparisons |
|
| 289 |
|
| 290 |
## Current Formula Status
|
| 291 |
|
|
|
|
| 334 |
| TinyStories held-out model | supports pressure dependence on model size |
|
| 335 |
| TinyStories streaming, 5 seeds | interaction has best mean final loss; interaction beats best static in 5/5 paired final-loss comparisons |
|
| 336 |
| OpenWebText10K streaming, 5 seeds | interaction decay has best mean final loss; top decay schedules beat best static in 5/5 paired comparisons |
|
| 337 |
+
| WikiText-103 streaming, 5 seeds | formula-derived L12 decay has best mean final loss; beats best static in 5/5 paired comparisons |
|
| 338 |
| cross-regime raw coefficient transfer | weaker than within-regime fit; supports regime-specific coefficients rather than universal numeric coefficients |
|
| 339 |
|
| 340 |
Latest TinyStories 5-seed streaming final-loss table:
|
|
|
|
| 402 |
`0.0473` and wins every paired seed comparison. This promotes OpenWebText10K
|
| 403 |
from exploratory support to a second multi-seed streaming validation regime.
|
| 404 |
|
| 405 |
+
Latest WikiText-103 5-seed streaming final-loss table:
|
| 406 |
+
|
| 407 |
+
| Condition | Mean final 4M validation loss | Std |
|
| 408 |
+
|---|---:|---:|
|
| 409 |
+
| `wikitext103_formula_l12` decay | 4.0808 | 0.0195 |
|
| 410 |
+
| `wikitext103_probe_blend` decay | 4.0961 | 0.0145 |
|
| 411 |
+
| `wikitext103_low_decay` decay | 4.1020 | 0.0166 |
|
| 412 |
+
| static `0.10` | 4.1105 | 0.0188 |
|
| 413 |
+
| static `0.08` | 4.1116 | 0.0186 |
|
| 414 |
+
| static `0.06` | 4.1197 | 0.0082 |
|
| 415 |
+
| static `0.14` | 4.1221 | 0.0155 |
|
| 416 |
+
| static `0.18` | 4.1304 | 0.0130 |
|
| 417 |
+
| static `0.04` | 4.1331 | 0.0227 |
|
| 418 |
+
| static `0.20` | 4.1394 | 0.0167 |
|
| 419 |
+
| static `0.02` | 4.1459 | 0.0165 |
|
| 420 |
+
| static `0.26` | 4.1784 | 0.0145 |
|
| 421 |
+
| static `0.00` | 4.1835 | 0.0165 |
|
| 422 |
+
| static `0.30` | 4.1946 | 0.0141 |
|
| 423 |
+
|
| 424 |
+
Paired final-loss result:
|
| 425 |
+
|
| 426 |
+
| Decay schedule | Paired wins vs best static |
|
| 427 |
+
|---|---:|
|
| 428 |
+
| `wikitext103_formula_l12` | 5/5 |
|
| 429 |
+
| `wikitext103_probe_blend` | 4/5 |
|
| 430 |
+
| `wikitext103_low_decay` | 4/5 |
|
| 431 |
+
|
| 432 |
+
The best static baseline in the clean WikiText-103 run is static dropout
|
| 433 |
+
`0.10` by mean final loss. The formula-derived L12 decay improves mean final
|
| 434 |
+
validation loss by about `0.0297` and wins every paired seed comparison. This
|
| 435 |
+
promotes WikiText-103 to a third multi-seed streaming validation regime.
|
| 436 |
+
|
| 437 |
## Completed Static Backtest Gate
|
| 438 |
|
| 439 |
The first offline coefficient backtest is complete. It is retained as supporting
|
|
|
|
| 456 |
|
| 457 |
## Immediate Next Action
|
| 458 |
|
| 459 |
+
Reconcile the TinyStories, OpenWebText10K, and WikiText-103 five-seed streaming
|
| 460 |
+
reports into the paper outline. The strongest current claim is now supported in
|
| 461 |
+
three regimes: formula-derived or regime-fitted decay schedules beat the best
|
| 462 |
+
static dropout baseline in paired five-seed final-loss comparisons.
|
| 463 |
+
|
| 464 |
+
The next empirical weakness is no longer a missing third text regime. The next
|
| 465 |
+
useful strengthening step is to test robustness across a controlled architecture
|
| 466 |
+
or token-budget change inside one established corpus regime, while preserving
|
| 467 |
+
the same MPS-only, five-seed validation standard.
|
| 468 |
|
| 469 |
## Next Training After Current Gate
|
| 470 |
|
|
|
|
| 475 |
```text
|
| 476 |
completed: TinyStories 5-seed streaming report
|
| 477 |
completed: OpenWebText10K 5-seed clean streaming report
|
| 478 |
+
completed: WikiText-103 5-seed clean streaming report
|
| 479 |
+
next: reconcile three-regime evidence into the paper, then choose one narrowed
|
| 480 |
+
robustness test
|
| 481 |
+
avoid: broad new sweep before three-regime report reconciliation
|
| 482 |
```
|
| 483 |
|
| 484 |
Evaluate with paired seed comparisons:
|
|
|
|
| 491 |
rank consistency across seeds
|
| 492 |
```
|
| 493 |
|
| 494 |
+
Because TinyStories, OpenWebText10K, and WikiText-103 decays win across paired
|
| 495 |
+
seeds, promote the cross-regime streaming claim to "supported in three regimes."
|
| 496 |
+
Do not yet claim universal numeric coefficients. The current defensible
|
| 497 |
+
paper-level claim is that the pressure-law structure and regime-specific fitting
|
| 498 |
+
procedure can produce dropout schedules that beat the best static dropout
|
| 499 |
+
baseline across multiple text regimes.
|
| 500 |
|
| 501 |
Latest streaming report:
|
| 502 |
|
| 503 |
```text
|
| 504 |
docs/tinystories_streaming_report.md
|
| 505 |
docs/openwebtext10k_streaming_report.md
|
| 506 |
+
docs/wikitext103_streaming_report.md
|
| 507 |
runs/streaming_tinystories_multiseed_validation_l12/combined_5seed_summary/
|
| 508 |
runs/openwebtext10k_streaming_report/l16_updated_formula_clean_5seed/
|
| 509 |
+
runs/wikitext103_streaming_report/l12_validation_5seed/
|
| 510 |
```
|
docs/wikitext103_streaming_report.md
ADDED
|
@@ -0,0 +1,198 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# WikiText-103 Streaming Validation
|
| 2 |
+
|
| 3 |
+
Date: 2026-05-31
|
| 4 |
+
|
| 5 |
+
This report combines 5 random seeds (1, 2, 3, 4, 5) from saved streaming runs.
|
| 6 |
+
No additional training is performed by this script; it reads saved
|
| 7 |
+
`metrics.jsonl` files.
|
| 8 |
+
|
| 9 |
+
Regime: WikiText-103 cached-corpus streaming setup with L12_H8_D320, 17,367,040 parameters, five prefixes from 250k to 4M tokens, and 1,000 optimizer steps per stage. This is a clean five-seed run including three dropout decay schedules and broad static dropout baselines from 0.00 through 0.30.
|
| 10 |
+
|
| 11 |
+
## Sources
|
| 12 |
+
|
| 13 |
+
- `runs/wikitext103_l12_streaming_validation_5seed/locked_stream/20260531-093525/metrics.jsonl`
|
| 14 |
+
|
| 15 |
+
## Condition Ranking By Final Loss
|
| 16 |
+
|
| 17 |
+
| Condition | Kind | N | Mean trajectory val | Std trajectory val | Mean final val | Std final val | Mean final gap | Dropout path |
|
| 18 |
+
|---|---|---:|---:|---:|---:|---:|---:|---|
|
| 19 |
+
| `wikitext103_formula_l12` | `anchor_decay` | 5 | 4.5711 | 0.0045 | 4.0808 | 0.0195 | 0.2817 | `0.30 -> 0.26 -> 0.18 -> 0.09 -> 0.02` |
|
| 20 |
+
| `wikitext103_probe_blend` | `anchor_decay` | 5 | 4.5635 | 0.0046 | 4.0961 | 0.0145 | 0.3287 | `0.19 -> 0.14 -> 0.09 -> 0.04 -> 0.01` |
|
| 21 |
+
| `wikitext103_low_decay` | `anchor_decay` | 5 | 4.5681 | 0.0073 | 4.1020 | 0.0166 | 0.3251 | `0.14 -> 0.14 -> 0.10 -> 0.06 -> 0.02` |
|
| 22 |
+
| `static_dropout_0.1` | `static` | 5 | 4.5836 | 0.0062 | 4.1105 | 0.0188 | 0.2687 | `0.10 -> 0.10 -> 0.10 -> 0.10 -> 0.10` |
|
| 23 |
+
| `static_dropout_0.08` | `static` | 5 | 4.5967 | 0.0073 | 4.1116 | 0.0186 | 0.2848 | `0.08 -> 0.08 -> 0.08 -> 0.08 -> 0.08` |
|
| 24 |
+
| `static_dropout_0.06` | `static` | 5 | 4.6186 | 0.0048 | 4.1197 | 0.0082 | 0.3131 | `0.06 -> 0.06 -> 0.06 -> 0.06 -> 0.06` |
|
| 25 |
+
| `static_dropout_0.14` | `static` | 5 | 4.5735 | 0.0077 | 4.1221 | 0.0155 | 0.2548 | `0.14 -> 0.14 -> 0.14 -> 0.14 -> 0.14` |
|
| 26 |
+
| `static_dropout_0.18` | `static` | 5 | 4.5756 | 0.0041 | 4.1304 | 0.0130 | 0.2289 | `0.18 -> 0.18 -> 0.18 -> 0.18 -> 0.18` |
|
| 27 |
+
| `static_dropout_0.04` | `static` | 5 | 4.6501 | 0.0077 | 4.1331 | 0.0227 | 0.3353 | `0.04 -> 0.04 -> 0.04 -> 0.04 -> 0.04` |
|
| 28 |
+
| `static_dropout_0.2` | `static` | 5 | 4.5794 | 0.0050 | 4.1394 | 0.0167 | 0.2239 | `0.20 -> 0.20 -> 0.20 -> 0.20 -> 0.20` |
|
| 29 |
+
| `static_dropout_0.02` | `static` | 5 | 4.6954 | 0.0086 | 4.1459 | 0.0165 | 0.3700 | `0.02 -> 0.02 -> 0.02 -> 0.02 -> 0.02` |
|
| 30 |
+
| `static_dropout_0.26` | `static` | 5 | 4.6063 | 0.0051 | 4.1784 | 0.0145 | 0.2008 | `0.26 -> 0.26 -> 0.26 -> 0.26 -> 0.26` |
|
| 31 |
+
| `static_dropout_0` | `static` | 5 | 4.7762 | 0.0109 | 4.1835 | 0.0165 | 0.4085 | `0.00 -> 0.00 -> 0.00 -> 0.00 -> 0.00` |
|
| 32 |
+
| `static_dropout_0.3` | `static` | 5 | 4.6253 | 0.0034 | 4.1946 | 0.0141 | 0.1819 | `0.30 -> 0.30 -> 0.30 -> 0.30 -> 0.30` |
|
| 33 |
+
|
| 34 |
+
## Paired Final-Loss Deltas
|
| 35 |
+
|
| 36 |
+
Negative `delta_vs_best_static` means the condition beat the best static
|
| 37 |
+
baseline for that seed.
|
| 38 |
+
|
| 39 |
+
| Seed | Condition | Final val | Best static | Best static final val | Delta vs best static |
|
| 40 |
+
|---:|---|---:|---|---:|---:|
|
| 41 |
+
| 1 | `wikitext103_formula_l12` | 4.0623 | `static_dropout_0.1` | 4.0807 | -0.0184 |
|
| 42 |
+
| 1 | `wikitext103_probe_blend` | 4.0738 | `static_dropout_0.1` | 4.0807 | -0.0069 |
|
| 43 |
+
| 1 | `wikitext103_low_decay` | 4.0854 | `static_dropout_0.1` | 4.0807 | +0.0047 |
|
| 44 |
+
| 1 | `static_dropout_0.1` | 4.0807 | `static_dropout_0.1` | 4.0807 | +0.0000 |
|
| 45 |
+
| 1 | `static_dropout_0.08` | 4.0893 | `static_dropout_0.1` | 4.0807 | +0.0086 |
|
| 46 |
+
| 1 | `static_dropout_0.06` | 4.1112 | `static_dropout_0.1` | 4.0807 | +0.0305 |
|
| 47 |
+
| 1 | `static_dropout_0.14` | 4.1082 | `static_dropout_0.1` | 4.0807 | +0.0275 |
|
| 48 |
+
| 1 | `static_dropout_0.18` | 4.1108 | `static_dropout_0.1` | 4.0807 | +0.0301 |
|
| 49 |
+
| 1 | `static_dropout_0.2` | 4.1162 | `static_dropout_0.1` | 4.0807 | +0.0355 |
|
| 50 |
+
| 1 | `static_dropout_0.04` | 4.1031 | `static_dropout_0.1` | 4.0807 | +0.0224 |
|
| 51 |
+
| 1 | `static_dropout_0.02` | 4.1371 | `static_dropout_0.1` | 4.0807 | +0.0564 |
|
| 52 |
+
| 1 | `static_dropout_0` | 4.1600 | `static_dropout_0.1` | 4.0807 | +0.0793 |
|
| 53 |
+
| 1 | `static_dropout_0.26` | 4.1557 | `static_dropout_0.1` | 4.0807 | +0.0750 |
|
| 54 |
+
| 1 | `static_dropout_0.3` | 4.1802 | `static_dropout_0.1` | 4.0807 | +0.0994 |
|
| 55 |
+
| 2 | `wikitext103_formula_l12` | 4.1123 | `static_dropout_0.06` | 4.1304 | -0.0181 |
|
| 56 |
+
| 2 | `wikitext103_probe_blend` | 4.1113 | `static_dropout_0.06` | 4.1304 | -0.0191 |
|
| 57 |
+
| 2 | `wikitext103_low_decay` | 4.1291 | `static_dropout_0.06` | 4.1304 | -0.0013 |
|
| 58 |
+
| 2 | `static_dropout_0.1` | 4.1320 | `static_dropout_0.06` | 4.1304 | +0.0016 |
|
| 59 |
+
| 2 | `static_dropout_0.08` | 4.1374 | `static_dropout_0.06` | 4.1304 | +0.0071 |
|
| 60 |
+
| 2 | `static_dropout_0.06` | 4.1304 | `static_dropout_0.06` | 4.1304 | +0.0000 |
|
| 61 |
+
| 2 | `static_dropout_0.14` | 4.1476 | `static_dropout_0.06` | 4.1304 | +0.0172 |
|
| 62 |
+
| 2 | `static_dropout_0.18` | 4.1471 | `static_dropout_0.06` | 4.1304 | +0.0167 |
|
| 63 |
+
| 2 | `static_dropout_0.2` | 4.1633 | `static_dropout_0.06` | 4.1304 | +0.0329 |
|
| 64 |
+
| 2 | `static_dropout_0.04` | 4.1648 | `static_dropout_0.06` | 4.1304 | +0.0344 |
|
| 65 |
+
| 2 | `static_dropout_0.02` | 4.1746 | `static_dropout_0.06` | 4.1304 | +0.0442 |
|
| 66 |
+
| 2 | `static_dropout_0` | 4.2030 | `static_dropout_0.06` | 4.1304 | +0.0726 |
|
| 67 |
+
| 2 | `static_dropout_0.26` | 4.1961 | `static_dropout_0.06` | 4.1304 | +0.0658 |
|
| 68 |
+
| 2 | `static_dropout_0.3` | 4.2155 | `static_dropout_0.06` | 4.1304 | +0.0852 |
|
| 69 |
+
| 3 | `wikitext103_formula_l12` | 4.0763 | `static_dropout_0.08` | 4.1036 | -0.0272 |
|
| 70 |
+
| 3 | `wikitext103_probe_blend` | 4.0934 | `static_dropout_0.08` | 4.1036 | -0.0102 |
|
| 71 |
+
| 3 | `wikitext103_low_decay` | 4.1006 | `static_dropout_0.08` | 4.1036 | -0.0030 |
|
| 72 |
+
| 3 | `static_dropout_0.1` | 4.1115 | `static_dropout_0.08` | 4.1036 | +0.0079 |
|
| 73 |
+
| 3 | `static_dropout_0.08` | 4.1036 | `static_dropout_0.08` | 4.1036 | +0.0000 |
|
| 74 |
+
| 3 | `static_dropout_0.06` | 4.1127 | `static_dropout_0.08` | 4.1036 | +0.0092 |
|
| 75 |
+
| 3 | `static_dropout_0.14` | 4.1240 | `static_dropout_0.08` | 4.1036 | +0.0204 |
|
| 76 |
+
| 3 | `static_dropout_0.18` | 4.1285 | `static_dropout_0.08` | 4.1036 | +0.0250 |
|
| 77 |
+
| 3 | `static_dropout_0.2` | 4.1367 | `static_dropout_0.08` | 4.1036 | +0.0332 |
|
| 78 |
+
| 3 | `static_dropout_0.04` | 4.1246 | `static_dropout_0.08` | 4.1036 | +0.0211 |
|
| 79 |
+
| 3 | `static_dropout_0.02` | 4.1443 | `static_dropout_0.08` | 4.1036 | +0.0408 |
|
| 80 |
+
| 3 | `static_dropout_0` | 4.1758 | `static_dropout_0.08` | 4.1036 | +0.0722 |
|
| 81 |
+
| 3 | `static_dropout_0.26` | 4.1796 | `static_dropout_0.08` | 4.1036 | +0.0761 |
|
| 82 |
+
| 3 | `static_dropout_0.3` | 4.1926 | `static_dropout_0.08` | 4.1036 | +0.0890 |
|
| 83 |
+
| 4 | `wikitext103_formula_l12` | 4.0845 | `static_dropout_0.1` | 4.1096 | -0.0251 |
|
| 84 |
+
| 4 | `wikitext103_probe_blend` | 4.0954 | `static_dropout_0.1` | 4.1096 | -0.0141 |
|
| 85 |
+
| 4 | `wikitext103_low_decay` | 4.0928 | `static_dropout_0.1` | 4.1096 | -0.0167 |
|
| 86 |
+
| 4 | `static_dropout_0.1` | 4.1096 | `static_dropout_0.1` | 4.1096 | +0.0000 |
|
| 87 |
+
| 4 | `static_dropout_0.08` | 4.1223 | `static_dropout_0.1` | 4.1096 | +0.0127 |
|
| 88 |
+
| 4 | `static_dropout_0.06` | 4.1188 | `static_dropout_0.1` | 4.1096 | +0.0093 |
|
| 89 |
+
| 4 | `static_dropout_0.14` | 4.1117 | `static_dropout_0.1` | 4.1096 | +0.0021 |
|
| 90 |
+
| 4 | `static_dropout_0.18` | 4.1330 | `static_dropout_0.1` | 4.1096 | +0.0234 |
|
| 91 |
+
| 4 | `static_dropout_0.2` | 4.1388 | `static_dropout_0.1` | 4.1096 | +0.0292 |
|
| 92 |
+
| 4 | `static_dropout_0.04` | 4.1312 | `static_dropout_0.1` | 4.1096 | +0.0217 |
|
| 93 |
+
| 4 | `static_dropout_0.02` | 4.1387 | `static_dropout_0.1` | 4.1096 | +0.0291 |
|
| 94 |
+
| 4 | `static_dropout_0` | 4.1853 | `static_dropout_0.1` | 4.1096 | +0.0757 |
|
| 95 |
+
| 4 | `static_dropout_0.26` | 4.1782 | `static_dropout_0.1` | 4.1096 | +0.0686 |
|
| 96 |
+
| 4 | `static_dropout_0.3` | 4.2007 | `static_dropout_0.1` | 4.1096 | +0.0912 |
|
| 97 |
+
| 5 | `wikitext103_formula_l12` | 4.0686 | `static_dropout_0.08` | 4.1056 | -0.0370 |
|
| 98 |
+
| 5 | `wikitext103_probe_blend` | 4.1066 | `static_dropout_0.08` | 4.1056 | +0.0009 |
|
| 99 |
+
| 5 | `wikitext103_low_decay` | 4.1021 | `static_dropout_0.08` | 4.1056 | -0.0035 |
|
| 100 |
+
| 5 | `static_dropout_0.1` | 4.1186 | `static_dropout_0.08` | 4.1056 | +0.0129 |
|
| 101 |
+
| 5 | `static_dropout_0.08` | 4.1056 | `static_dropout_0.08` | 4.1056 | +0.0000 |
|
| 102 |
+
| 5 | `static_dropout_0.06` | 4.1253 | `static_dropout_0.08` | 4.1056 | +0.0197 |
|
| 103 |
+
| 5 | `static_dropout_0.14` | 4.1192 | `static_dropout_0.08` | 4.1056 | +0.0135 |
|
| 104 |
+
| 5 | `static_dropout_0.18` | 4.1325 | `static_dropout_0.08` | 4.1056 | +0.0269 |
|
| 105 |
+
| 5 | `static_dropout_0.2` | 4.1418 | `static_dropout_0.08` | 4.1056 | +0.0362 |
|
| 106 |
+
| 5 | `static_dropout_0.04` | 4.1419 | `static_dropout_0.08` | 4.1056 | +0.0363 |
|
| 107 |
+
| 5 | `static_dropout_0.02` | 4.1346 | `static_dropout_0.08` | 4.1056 | +0.0290 |
|
| 108 |
+
| 5 | `static_dropout_0` | 4.1934 | `static_dropout_0.08` | 4.1056 | +0.0878 |
|
| 109 |
+
| 5 | `static_dropout_0.26` | 4.1821 | `static_dropout_0.08` | 4.1056 | +0.0765 |
|
| 110 |
+
| 5 | `static_dropout_0.3` | 4.1841 | `static_dropout_0.08` | 4.1056 | +0.0785 |
|
| 111 |
+
|
| 112 |
+
## Stage Trajectory
|
| 113 |
+
|
| 114 |
+
| Stage | Prefix tokens | Condition | Dropout | N | Mean val | Std val | Mean train | Mean gap |
|
| 115 |
+
|---:|---:|---|---:|---:|---:|---:|---:|---:|
|
| 116 |
+
| 0 | 250,000 | `static_dropout_0.18` | 0.180 | 5 | 5.1616 | 0.0150 | 3.9964 | 1.1652 |
|
| 117 |
+
| 0 | 250,000 | `wikitext103_low_decay` | 0.140 | 5 | 5.1635 | 0.0220 | 3.9051 | 1.2585 |
|
| 118 |
+
| 0 | 250,000 | `static_dropout_0.14` | 0.140 | 5 | 5.1635 | 0.0220 | 3.9051 | 1.2585 |
|
| 119 |
+
| 0 | 250,000 | `wikitext103_probe_blend` | 0.190 | 5 | 5.1659 | 0.0171 | 4.0201 | 1.1458 |
|
| 120 |
+
| 0 | 250,000 | `static_dropout_0.1` | 0.100 | 5 | 5.1699 | 0.0237 | 3.8219 | 1.3480 |
|
| 121 |
+
| 0 | 250,000 | `static_dropout_0.2` | 0.200 | 5 | 5.1701 | 0.0141 | 4.0363 | 1.1338 |
|
| 122 |
+
| 0 | 250,000 | `static_dropout_0.08` | 0.080 | 5 | 5.1894 | 0.0161 | 3.7619 | 1.4274 |
|
| 123 |
+
| 0 | 250,000 | `static_dropout_0.26` | 0.260 | 5 | 5.1940 | 0.0161 | 4.1496 | 1.0444 |
|
| 124 |
+
| 0 | 250,000 | `wikitext103_formula_l12` | 0.300 | 5 | 5.2148 | 0.0181 | 4.2131 | 1.0017 |
|
| 125 |
+
| 0 | 250,000 | `static_dropout_0.3` | 0.300 | 5 | 5.2148 | 0.0181 | 4.2131 | 1.0017 |
|
| 126 |
+
| 0 | 250,000 | `static_dropout_0.06` | 0.060 | 5 | 5.2154 | 0.0173 | 3.7128 | 1.5026 |
|
| 127 |
+
| 0 | 250,000 | `static_dropout_0.04` | 0.040 | 5 | 5.2378 | 0.0186 | 3.6441 | 1.5938 |
|
| 128 |
+
| 0 | 250,000 | `static_dropout_0.02` | 0.020 | 5 | 5.2750 | 0.0255 | 3.5725 | 1.7025 |
|
| 129 |
+
| 0 | 250,000 | `static_dropout_0` | 0.000 | 5 | 5.3403 | 0.0270 | 3.5230 | 1.8172 |
|
| 130 |
+
| 1 | 500,000 | `wikitext103_probe_blend` | 0.140 | 5 | 4.7872 | 0.0269 | 3.6846 | 1.1027 |
|
| 131 |
+
| 1 | 500,000 | `static_dropout_0.2` | 0.200 | 5 | 4.7873 | 0.0236 | 3.7914 | 0.9959 |
|
| 132 |
+
| 1 | 500,000 | `static_dropout_0.18` | 0.180 | 5 | 4.7946 | 0.0206 | 3.7572 | 1.0375 |
|
| 133 |
+
| 1 | 500,000 | `static_dropout_0.14` | 0.140 | 5 | 4.8001 | 0.0198 | 3.6650 | 1.1351 |
|
| 134 |
+
| 1 | 500,000 | `wikitext103_low_decay` | 0.140 | 5 | 4.8001 | 0.0198 | 3.6650 | 1.1351 |
|
| 135 |
+
| 1 | 500,000 | `wikitext103_formula_l12` | 0.260 | 5 | 4.8053 | 0.0278 | 3.9182 | 0.8871 |
|
| 136 |
+
| 1 | 500,000 | `static_dropout_0.26` | 0.260 | 5 | 4.8081 | 0.0216 | 3.9053 | 0.9028 |
|
| 137 |
+
| 1 | 500,000 | `static_dropout_0.3` | 0.300 | 5 | 4.8242 | 0.0296 | 3.9765 | 0.8476 |
|
| 138 |
+
| 1 | 500,000 | `static_dropout_0.1` | 0.100 | 5 | 4.8332 | 0.0258 | 3.5637 | 1.2695 |
|
| 139 |
+
| 1 | 500,000 | `static_dropout_0.08` | 0.080 | 5 | 4.8576 | 0.0239 | 3.5036 | 1.3540 |
|
| 140 |
+
| 1 | 500,000 | `static_dropout_0.06` | 0.060 | 5 | 4.8947 | 0.0213 | 3.4394 | 1.4552 |
|
| 141 |
+
| 1 | 500,000 | `static_dropout_0.04` | 0.040 | 5 | 4.9573 | 0.0250 | 3.3515 | 1.6058 |
|
| 142 |
+
| 1 | 500,000 | `static_dropout_0.02` | 0.020 | 5 | 5.0451 | 0.0169 | 3.2612 | 1.7839 |
|
| 143 |
+
| 1 | 500,000 | `static_dropout_0` | 0.000 | 5 | 5.1741 | 0.0252 | 3.1506 | 2.0235 |
|
| 144 |
+
| 2 | 1,000,000 | `wikitext103_formula_l12` | 0.180 | 5 | 4.4938 | 0.0147 | 3.8283 | 0.6655 |
|
| 145 |
+
| 2 | 1,000,000 | `wikitext103_probe_blend` | 0.090 | 5 | 4.4940 | 0.0159 | 3.6495 | 0.8445 |
|
| 146 |
+
| 2 | 1,000,000 | `static_dropout_0.14` | 0.140 | 5 | 4.5001 | 0.0148 | 3.7163 | 0.7838 |
|
| 147 |
+
| 2 | 1,000,000 | `wikitext103_low_decay` | 0.100 | 5 | 4.5013 | 0.0158 | 3.6607 | 0.8406 |
|
| 148 |
+
| 2 | 1,000,000 | `static_dropout_0.18` | 0.180 | 5 | 4.5023 | 0.0185 | 3.7866 | 0.7157 |
|
| 149 |
+
| 2 | 1,000,000 | `static_dropout_0.2` | 0.200 | 5 | 4.5060 | 0.0204 | 3.8148 | 0.6913 |
|
| 150 |
+
| 2 | 1,000,000 | `static_dropout_0.1` | 0.100 | 5 | 4.5186 | 0.0135 | 3.6524 | 0.8662 |
|
| 151 |
+
| 2 | 1,000,000 | `static_dropout_0.26` | 0.260 | 5 | 4.5262 | 0.0101 | 3.9071 | 0.6191 |
|
| 152 |
+
| 2 | 1,000,000 | `static_dropout_0.08` | 0.080 | 5 | 4.5326 | 0.0117 | 3.6000 | 0.9326 |
|
| 153 |
+
| 2 | 1,000,000 | `static_dropout_0.3` | 0.300 | 5 | 4.5462 | 0.0127 | 3.9708 | 0.5754 |
|
| 154 |
+
| 2 | 1,000,000 | `static_dropout_0.06` | 0.060 | 5 | 4.5574 | 0.0126 | 3.5554 | 1.0020 |
|
| 155 |
+
| 2 | 1,000,000 | `static_dropout_0.04` | 0.040 | 5 | 4.5959 | 0.0146 | 3.5030 | 1.0929 |
|
| 156 |
+
| 2 | 1,000,000 | `static_dropout_0.02` | 0.020 | 5 | 4.6558 | 0.0159 | 3.4324 | 1.2234 |
|
| 157 |
+
| 2 | 1,000,000 | `static_dropout_0` | 0.000 | 5 | 4.7661 | 0.0324 | 3.3658 | 1.4003 |
|
| 158 |
+
| 3 | 2,000,000 | `wikitext103_formula_l12` | 0.090 | 5 | 4.2607 | 0.0181 | 3.8089 | 0.4518 |
|
| 159 |
+
| 3 | 2,000,000 | `wikitext103_low_decay` | 0.060 | 5 | 4.2736 | 0.0228 | 3.7474 | 0.5261 |
|
| 160 |
+
| 3 | 2,000,000 | `wikitext103_probe_blend` | 0.040 | 5 | 4.2743 | 0.0174 | 3.7200 | 0.5543 |
|
| 161 |
+
| 3 | 2,000,000 | `static_dropout_0.14` | 0.140 | 5 | 4.2816 | 0.0214 | 3.8287 | 0.4529 |
|
| 162 |
+
| 3 | 2,000,000 | `static_dropout_0.1` | 0.100 | 5 | 4.2857 | 0.0190 | 3.7809 | 0.5048 |
|
| 163 |
+
| 3 | 2,000,000 | `static_dropout_0.18` | 0.180 | 5 | 4.2889 | 0.0195 | 3.8768 | 0.4121 |
|
| 164 |
+
| 3 | 2,000,000 | `static_dropout_0.08` | 0.080 | 5 | 4.2925 | 0.0166 | 3.7519 | 0.5406 |
|
| 165 |
+
| 3 | 2,000,000 | `static_dropout_0.2` | 0.200 | 5 | 4.2942 | 0.0160 | 3.8982 | 0.3960 |
|
| 166 |
+
| 3 | 2,000,000 | `static_dropout_0.06` | 0.060 | 5 | 4.3059 | 0.0182 | 3.7318 | 0.5741 |
|
| 167 |
+
| 3 | 2,000,000 | `static_dropout_0.26` | 0.260 | 5 | 4.3248 | 0.0179 | 3.9655 | 0.3593 |
|
| 168 |
+
| 3 | 2,000,000 | `static_dropout_0.04` | 0.040 | 5 | 4.3262 | 0.0165 | 3.7038 | 0.6225 |
|
| 169 |
+
| 3 | 2,000,000 | `static_dropout_0.3` | 0.300 | 5 | 4.3467 | 0.0159 | 4.0097 | 0.3370 |
|
| 170 |
+
| 3 | 2,000,000 | `static_dropout_0.02` | 0.020 | 5 | 4.3551 | 0.0265 | 3.6673 | 0.6878 |
|
| 171 |
+
| 3 | 2,000,000 | `static_dropout_0` | 0.000 | 5 | 4.4174 | 0.0188 | 3.6409 | 0.7765 |
|
| 172 |
+
| 4 | 4,000,000 | `wikitext103_formula_l12` | 0.020 | 5 | 4.0808 | 0.0195 | 3.7991 | 0.2817 |
|
| 173 |
+
| 4 | 4,000,000 | `wikitext103_probe_blend` | 0.010 | 5 | 4.0961 | 0.0145 | 3.7674 | 0.3287 |
|
| 174 |
+
| 4 | 4,000,000 | `wikitext103_low_decay` | 0.020 | 5 | 4.1020 | 0.0166 | 3.7769 | 0.3251 |
|
| 175 |
+
| 4 | 4,000,000 | `static_dropout_0.1` | 0.100 | 5 | 4.1105 | 0.0188 | 3.8417 | 0.2687 |
|
| 176 |
+
| 4 | 4,000,000 | `static_dropout_0.08` | 0.080 | 5 | 4.1116 | 0.0186 | 3.8268 | 0.2848 |
|
| 177 |
+
| 4 | 4,000,000 | `static_dropout_0.06` | 0.060 | 5 | 4.1197 | 0.0082 | 3.8066 | 0.3131 |
|
| 178 |
+
| 4 | 4,000,000 | `static_dropout_0.14` | 0.140 | 5 | 4.1221 | 0.0155 | 3.8674 | 0.2548 |
|
| 179 |
+
| 4 | 4,000,000 | `static_dropout_0.18` | 0.180 | 5 | 4.1304 | 0.0130 | 3.9015 | 0.2289 |
|
| 180 |
+
| 4 | 4,000,000 | `static_dropout_0.04` | 0.040 | 5 | 4.1331 | 0.0227 | 3.7978 | 0.3353 |
|
| 181 |
+
| 4 | 4,000,000 | `static_dropout_0.2` | 0.200 | 5 | 4.1394 | 0.0167 | 3.9155 | 0.2239 |
|
| 182 |
+
| 4 | 4,000,000 | `static_dropout_0.02` | 0.020 | 5 | 4.1459 | 0.0165 | 3.7759 | 0.3700 |
|
| 183 |
+
| 4 | 4,000,000 | `static_dropout_0.26` | 0.260 | 5 | 4.1784 | 0.0145 | 3.9775 | 0.2008 |
|
| 184 |
+
| 4 | 4,000,000 | `static_dropout_0` | 0.000 | 5 | 4.1835 | 0.0165 | 3.7750 | 0.4085 |
|
| 185 |
+
| 4 | 4,000,000 | `static_dropout_0.3` | 0.300 | 5 | 4.1946 | 0.0141 | 4.0127 | 0.1819 |
|
| 186 |
+
|
| 187 |
+
## Interpretation
|
| 188 |
+
|
| 189 |
+
- `wikitext103_formula_l12` has the best 5-seed mean final validation loss: 4.0808 +/- 0.0195.
|
| 190 |
+
- The second-best final condition is `wikitext103_probe_blend` at 4.0961 +/- 0.0145.
|
| 191 |
+
- The best static baseline by mean final loss is `static_dropout_0.1` at 4.1105 +/- 0.0188.
|
| 192 |
+
- `wikitext103_formula_l12` beats the per-seed best static baseline in 5/5 seeds; worst paired delta is -0.0181.
|
| 193 |
+
- `wikitext103_probe_blend` beats the per-seed best static baseline in 4/5 seeds; worst paired delta is +0.0009.
|
| 194 |
+
- `wikitext103_low_decay` beats the per-seed best static baseline in 4/5 seeds; worst paired delta is +0.0047.
|
| 195 |
+
- The best first-stage condition is `static_dropout_0.18` at prefix 250,000 with mean validation loss 5.1616; compare this with the final ranking before claiming a schedule is uniformly better.
|
| 196 |
+
- This is a saved-run streaming validation artifact. Treat it as strong
|
| 197 |
+
evidence only when the tested conditions, seeds, static baselines, and
|
| 198 |
+
stream protocol match the claim being made.
|
runs/wikitext103_l12_streaming_validation_5seed/locked_stream/20260531-093525/RESULT_SUMMARY.md
ADDED
|
@@ -0,0 +1,122 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Locked Streaming Dropout Summary
|
| 2 |
+
|
| 3 |
+
Run directory: `runs/wikitext103_l12_streaming_validation_5seed/locked_stream/20260531-093525`
|
| 4 |
+
|
| 5 |
+
Model: `L12_H8_D320` causal Transformer, 17,367,040 parameters, 12 layers, 8 heads, 320 embedding dim.
|
| 6 |
+
Training per stage: 1,000 steps. Sampled tokens are cumulative in each stage row. Seeds present: 1, 2, 3, 4, 5.
|
| 7 |
+
|
| 8 |
+
## Condition Ranking
|
| 9 |
+
|
| 10 |
+
| Condition | Kind | Final dropout | Mean trajectory val loss | Final val loss | Final gap | Dropout path |
|
| 11 |
+
|---|---|---:|---:|---:|---:|---|
|
| 12 |
+
| `wikitext103_probe_blend` | anchor_decay | 0.01 | 4.5635 | 4.0961 | 0.3287 | 0.19 -> 0.14 -> 0.09 -> 0.04 -> 0.01 |
|
| 13 |
+
| `wikitext103_low_decay` | anchor_decay | 0.02 | 4.5681 | 4.1020 | 0.3251 | 0.14 -> 0.14 -> 0.10 -> 0.06 -> 0.02 |
|
| 14 |
+
| `wikitext103_formula_l12` | anchor_decay | 0.02 | 4.5711 | 4.0808 | 0.2817 | 0.30 -> 0.26 -> 0.18 -> 0.09 -> 0.02 |
|
| 15 |
+
| `static_dropout_0.14` | static | 0.14 | 4.5735 | 4.1221 | 0.2548 | 0.14 -> 0.14 -> 0.14 -> 0.14 -> 0.14 |
|
| 16 |
+
| `static_dropout_0.18` | static | 0.18 | 4.5756 | 4.1304 | 0.2289 | 0.18 -> 0.18 -> 0.18 -> 0.18 -> 0.18 |
|
| 17 |
+
| `static_dropout_0.2` | static | 0.20 | 4.5794 | 4.1394 | 0.2239 | 0.20 -> 0.20 -> 0.20 -> 0.20 -> 0.20 |
|
| 18 |
+
| `static_dropout_0.1` | static | 0.10 | 4.5836 | 4.1105 | 0.2687 | 0.10 -> 0.10 -> 0.10 -> 0.10 -> 0.10 |
|
| 19 |
+
| `static_dropout_0.08` | static | 0.08 | 4.5967 | 4.1116 | 0.2848 | 0.08 -> 0.08 -> 0.08 -> 0.08 -> 0.08 |
|
| 20 |
+
| `static_dropout_0.26` | static | 0.26 | 4.6063 | 4.1784 | 0.2008 | 0.26 -> 0.26 -> 0.26 -> 0.26 -> 0.26 |
|
| 21 |
+
| `static_dropout_0.06` | static | 0.06 | 4.6186 | 4.1197 | 0.3131 | 0.06 -> 0.06 -> 0.06 -> 0.06 -> 0.06 |
|
| 22 |
+
| `static_dropout_0.3` | static | 0.30 | 4.6253 | 4.1946 | 0.1819 | 0.30 -> 0.30 -> 0.30 -> 0.30 -> 0.30 |
|
| 23 |
+
| `static_dropout_0.04` | static | 0.04 | 4.6501 | 4.1331 | 0.3353 | 0.04 -> 0.04 -> 0.04 -> 0.04 -> 0.04 |
|
| 24 |
+
| `static_dropout_0.02` | static | 0.02 | 4.6954 | 4.1459 | 0.3700 | 0.02 -> 0.02 -> 0.02 -> 0.02 -> 0.02 |
|
| 25 |
+
| `static_dropout_0` | static | 0.00 | 4.7762 | 4.1835 | 0.4085 | 0.00 -> 0.00 -> 0.00 -> 0.00 -> 0.00 |
|
| 26 |
+
|
| 27 |
+
## Stage Trajectory
|
| 28 |
+
|
| 29 |
+
### Stage 0: 250,000 Prefix Tokens
|
| 30 |
+
|
| 31 |
+
| Condition | Dropout | Mean val loss | Mean train loss | Mean gap | N |
|
| 32 |
+
|---|---:|---:|---:|---:|---:|
|
| 33 |
+
| `static_dropout_0.18` | 0.18 | 5.1616 | 3.9964 | 1.1652 | 5 |
|
| 34 |
+
| `wikitext103_low_decay` | 0.14 | 5.1635 | 3.9051 | 1.2585 | 5 |
|
| 35 |
+
| `static_dropout_0.14` | 0.14 | 5.1635 | 3.9051 | 1.2585 | 5 |
|
| 36 |
+
| `wikitext103_probe_blend` | 0.19 | 5.1659 | 4.0201 | 1.1458 | 5 |
|
| 37 |
+
| `static_dropout_0.1` | 0.10 | 5.1699 | 3.8219 | 1.3480 | 5 |
|
| 38 |
+
| `static_dropout_0.2` | 0.20 | 5.1701 | 4.0363 | 1.1338 | 5 |
|
| 39 |
+
| `static_dropout_0.08` | 0.08 | 5.1894 | 3.7619 | 1.4274 | 5 |
|
| 40 |
+
| `static_dropout_0.26` | 0.26 | 5.1940 | 4.1496 | 1.0444 | 5 |
|
| 41 |
+
| `wikitext103_formula_l12` | 0.30 | 5.2148 | 4.2131 | 1.0017 | 5 |
|
| 42 |
+
| `static_dropout_0.3` | 0.30 | 5.2148 | 4.2131 | 1.0017 | 5 |
|
| 43 |
+
| `static_dropout_0.06` | 0.06 | 5.2154 | 3.7128 | 1.5026 | 5 |
|
| 44 |
+
| `static_dropout_0.04` | 0.04 | 5.2378 | 3.6441 | 1.5938 | 5 |
|
| 45 |
+
| `static_dropout_0.02` | 0.02 | 5.2750 | 3.5725 | 1.7025 | 5 |
|
| 46 |
+
| `static_dropout_0` | 0.00 | 5.3403 | 3.5230 | 1.8172 | 5 |
|
| 47 |
+
|
| 48 |
+
### Stage 1: 500,000 Prefix Tokens
|
| 49 |
+
|
| 50 |
+
| Condition | Dropout | Mean val loss | Mean train loss | Mean gap | N |
|
| 51 |
+
|---|---:|---:|---:|---:|---:|
|
| 52 |
+
| `wikitext103_probe_blend` | 0.14 | 4.7872 | 3.6846 | 1.1027 | 5 |
|
| 53 |
+
| `static_dropout_0.2` | 0.20 | 4.7873 | 3.7914 | 0.9959 | 5 |
|
| 54 |
+
| `static_dropout_0.18` | 0.18 | 4.7946 | 3.7572 | 1.0375 | 5 |
|
| 55 |
+
| `static_dropout_0.14` | 0.14 | 4.8001 | 3.6650 | 1.1351 | 5 |
|
| 56 |
+
| `wikitext103_low_decay` | 0.14 | 4.8001 | 3.6650 | 1.1351 | 5 |
|
| 57 |
+
| `wikitext103_formula_l12` | 0.26 | 4.8053 | 3.9182 | 0.8871 | 5 |
|
| 58 |
+
| `static_dropout_0.26` | 0.26 | 4.8081 | 3.9053 | 0.9028 | 5 |
|
| 59 |
+
| `static_dropout_0.3` | 0.30 | 4.8242 | 3.9765 | 0.8476 | 5 |
|
| 60 |
+
| `static_dropout_0.1` | 0.10 | 4.8332 | 3.5637 | 1.2695 | 5 |
|
| 61 |
+
| `static_dropout_0.08` | 0.08 | 4.8576 | 3.5036 | 1.3540 | 5 |
|
| 62 |
+
| `static_dropout_0.06` | 0.06 | 4.8947 | 3.4394 | 1.4552 | 5 |
|
| 63 |
+
| `static_dropout_0.04` | 0.04 | 4.9573 | 3.3515 | 1.6058 | 5 |
|
| 64 |
+
| `static_dropout_0.02` | 0.02 | 5.0451 | 3.2612 | 1.7839 | 5 |
|
| 65 |
+
| `static_dropout_0` | 0.00 | 5.1741 | 3.1506 | 2.0235 | 5 |
|
| 66 |
+
|
| 67 |
+
### Stage 2: 1,000,000 Prefix Tokens
|
| 68 |
+
|
| 69 |
+
| Condition | Dropout | Mean val loss | Mean train loss | Mean gap | N |
|
| 70 |
+
|---|---:|---:|---:|---:|---:|
|
| 71 |
+
| `wikitext103_formula_l12` | 0.18 | 4.4938 | 3.8283 | 0.6655 | 5 |
|
| 72 |
+
| `wikitext103_probe_blend` | 0.09 | 4.4940 | 3.6495 | 0.8445 | 5 |
|
| 73 |
+
| `static_dropout_0.14` | 0.14 | 4.5001 | 3.7163 | 0.7838 | 5 |
|
| 74 |
+
| `wikitext103_low_decay` | 0.10 | 4.5013 | 3.6607 | 0.8406 | 5 |
|
| 75 |
+
| `static_dropout_0.18` | 0.18 | 4.5023 | 3.7866 | 0.7157 | 5 |
|
| 76 |
+
| `static_dropout_0.2` | 0.20 | 4.5060 | 3.8148 | 0.6913 | 5 |
|
| 77 |
+
| `static_dropout_0.1` | 0.10 | 4.5186 | 3.6524 | 0.8662 | 5 |
|
| 78 |
+
| `static_dropout_0.26` | 0.26 | 4.5262 | 3.9071 | 0.6191 | 5 |
|
| 79 |
+
| `static_dropout_0.08` | 0.08 | 4.5326 | 3.6000 | 0.9326 | 5 |
|
| 80 |
+
| `static_dropout_0.3` | 0.30 | 4.5462 | 3.9708 | 0.5754 | 5 |
|
| 81 |
+
| `static_dropout_0.06` | 0.06 | 4.5574 | 3.5554 | 1.0020 | 5 |
|
| 82 |
+
| `static_dropout_0.04` | 0.04 | 4.5959 | 3.5030 | 1.0929 | 5 |
|
| 83 |
+
| `static_dropout_0.02` | 0.02 | 4.6558 | 3.4324 | 1.2234 | 5 |
|
| 84 |
+
| `static_dropout_0` | 0.00 | 4.7661 | 3.3658 | 1.4003 | 5 |
|
| 85 |
+
|
| 86 |
+
### Stage 3: 2,000,000 Prefix Tokens
|
| 87 |
+
|
| 88 |
+
| Condition | Dropout | Mean val loss | Mean train loss | Mean gap | N |
|
| 89 |
+
|---|---:|---:|---:|---:|---:|
|
| 90 |
+
| `wikitext103_formula_l12` | 0.09 | 4.2607 | 3.8089 | 0.4518 | 5 |
|
| 91 |
+
| `wikitext103_low_decay` | 0.06 | 4.2736 | 3.7474 | 0.5261 | 5 |
|
| 92 |
+
| `wikitext103_probe_blend` | 0.04 | 4.2743 | 3.7200 | 0.5543 | 5 |
|
| 93 |
+
| `static_dropout_0.14` | 0.14 | 4.2816 | 3.8287 | 0.4529 | 5 |
|
| 94 |
+
| `static_dropout_0.1` | 0.10 | 4.2857 | 3.7809 | 0.5048 | 5 |
|
| 95 |
+
| `static_dropout_0.18` | 0.18 | 4.2889 | 3.8768 | 0.4121 | 5 |
|
| 96 |
+
| `static_dropout_0.08` | 0.08 | 4.2925 | 3.7519 | 0.5406 | 5 |
|
| 97 |
+
| `static_dropout_0.2` | 0.20 | 4.2942 | 3.8982 | 0.3960 | 5 |
|
| 98 |
+
| `static_dropout_0.06` | 0.06 | 4.3059 | 3.7318 | 0.5741 | 5 |
|
| 99 |
+
| `static_dropout_0.26` | 0.26 | 4.3248 | 3.9655 | 0.3593 | 5 |
|
| 100 |
+
| `static_dropout_0.04` | 0.04 | 4.3262 | 3.7038 | 0.6225 | 5 |
|
| 101 |
+
| `static_dropout_0.3` | 0.30 | 4.3467 | 4.0097 | 0.3370 | 5 |
|
| 102 |
+
| `static_dropout_0.02` | 0.02 | 4.3551 | 3.6673 | 0.6878 | 5 |
|
| 103 |
+
| `static_dropout_0` | 0.00 | 4.4174 | 3.6409 | 0.7765 | 5 |
|
| 104 |
+
|
| 105 |
+
### Stage 4: 4,000,000 Prefix Tokens
|
| 106 |
+
|
| 107 |
+
| Condition | Dropout | Mean val loss | Mean train loss | Mean gap | N |
|
| 108 |
+
|---|---:|---:|---:|---:|---:|
|
| 109 |
+
| `wikitext103_formula_l12` | 0.02 | 4.0808 | 3.7991 | 0.2817 | 5 |
|
| 110 |
+
| `wikitext103_probe_blend` | 0.01 | 4.0961 | 3.7674 | 0.3287 | 5 |
|
| 111 |
+
| `wikitext103_low_decay` | 0.02 | 4.1020 | 3.7769 | 0.3251 | 5 |
|
| 112 |
+
| `static_dropout_0.1` | 0.10 | 4.1105 | 3.8417 | 0.2687 | 5 |
|
| 113 |
+
| `static_dropout_0.08` | 0.08 | 4.1116 | 3.8268 | 0.2848 | 5 |
|
| 114 |
+
| `static_dropout_0.06` | 0.06 | 4.1197 | 3.8066 | 0.3131 | 5 |
|
| 115 |
+
| `static_dropout_0.14` | 0.14 | 4.1221 | 3.8674 | 0.2548 | 5 |
|
| 116 |
+
| `static_dropout_0.18` | 0.18 | 4.1304 | 3.9015 | 0.2289 | 5 |
|
| 117 |
+
| `static_dropout_0.04` | 0.04 | 4.1331 | 3.7978 | 0.3353 | 5 |
|
| 118 |
+
| `static_dropout_0.2` | 0.20 | 4.1394 | 3.9155 | 0.2239 | 5 |
|
| 119 |
+
| `static_dropout_0.02` | 0.02 | 4.1459 | 3.7759 | 0.3700 | 5 |
|
| 120 |
+
| `static_dropout_0.26` | 0.26 | 4.1784 | 3.9775 | 0.2008 | 5 |
|
| 121 |
+
| `static_dropout_0` | 0.00 | 4.1835 | 3.7750 | 0.4085 | 5 |
|
| 122 |
+
| `static_dropout_0.3` | 0.30 | 4.1946 | 4.0127 | 0.1819 | 5 |
|
runs/wikitext103_l12_streaming_validation_5seed/locked_stream/20260531-093525/config.json
ADDED
|
@@ -0,0 +1,199 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"args": {
|
| 3 |
+
"mode": "locked_stream",
|
| 4 |
+
"corpus": null,
|
| 5 |
+
"corpus_glob": null,
|
| 6 |
+
"text_column": "text",
|
| 7 |
+
"use_cached_data": true,
|
| 8 |
+
"output_dir": "runs/wikitext103_l12_streaming_validation_5seed",
|
| 9 |
+
"resume_from": null,
|
| 10 |
+
"cache_dir": ".cache/dropout_decay_wikitext103",
|
| 11 |
+
"models": [
|
| 12 |
+
"L12_H8_D320=12x8x320"
|
| 13 |
+
],
|
| 14 |
+
"seeds": [
|
| 15 |
+
1,
|
| 16 |
+
2,
|
| 17 |
+
3,
|
| 18 |
+
4,
|
| 19 |
+
5
|
| 20 |
+
],
|
| 21 |
+
"token_limits": [
|
| 22 |
+
5000000
|
| 23 |
+
],
|
| 24 |
+
"stream_token_caps": [
|
| 25 |
+
250000,
|
| 26 |
+
500000,
|
| 27 |
+
1000000,
|
| 28 |
+
2000000,
|
| 29 |
+
4000000
|
| 30 |
+
],
|
| 31 |
+
"val_tokens": 500000,
|
| 32 |
+
"allow_short_corpus": false,
|
| 33 |
+
"force_retokenize": false,
|
| 34 |
+
"vocab_size": 4096,
|
| 35 |
+
"tokenizer_train_chars": 10000000,
|
| 36 |
+
"block_size": 128,
|
| 37 |
+
"batch_size": 16,
|
| 38 |
+
"steps": 2000,
|
| 39 |
+
"stage_steps": 1000,
|
| 40 |
+
"dropout_rates": [
|
| 41 |
+
0.0,
|
| 42 |
+
0.02,
|
| 43 |
+
0.04,
|
| 44 |
+
0.06,
|
| 45 |
+
0.08,
|
| 46 |
+
0.1,
|
| 47 |
+
0.14,
|
| 48 |
+
0.18,
|
| 49 |
+
0.2,
|
| 50 |
+
0.26,
|
| 51 |
+
0.3
|
| 52 |
+
],
|
| 53 |
+
"decays": [],
|
| 54 |
+
"anchor_decays": [
|
| 55 |
+
{
|
| 56 |
+
"name": "wikitext103_formula_l12",
|
| 57 |
+
"kind": "anchor_decay",
|
| 58 |
+
"initial": 0.3,
|
| 59 |
+
"final": 0.02,
|
| 60 |
+
"schedule": "log_prefix_anchor",
|
| 61 |
+
"decay_tokens": null,
|
| 62 |
+
"anchors": [
|
| 63 |
+
[
|
| 64 |
+
250000,
|
| 65 |
+
0.3
|
| 66 |
+
],
|
| 67 |
+
[
|
| 68 |
+
500000,
|
| 69 |
+
0.26
|
| 70 |
+
],
|
| 71 |
+
[
|
| 72 |
+
1000000,
|
| 73 |
+
0.18
|
| 74 |
+
],
|
| 75 |
+
[
|
| 76 |
+
2000000,
|
| 77 |
+
0.09
|
| 78 |
+
],
|
| 79 |
+
[
|
| 80 |
+
4000000,
|
| 81 |
+
0.02
|
| 82 |
+
]
|
| 83 |
+
]
|
| 84 |
+
},
|
| 85 |
+
{
|
| 86 |
+
"name": "wikitext103_probe_blend",
|
| 87 |
+
"kind": "anchor_decay",
|
| 88 |
+
"initial": 0.19,
|
| 89 |
+
"final": 0.01,
|
| 90 |
+
"schedule": "log_prefix_anchor",
|
| 91 |
+
"decay_tokens": null,
|
| 92 |
+
"anchors": [
|
| 93 |
+
[
|
| 94 |
+
250000,
|
| 95 |
+
0.19
|
| 96 |
+
],
|
| 97 |
+
[
|
| 98 |
+
500000,
|
| 99 |
+
0.14
|
| 100 |
+
],
|
| 101 |
+
[
|
| 102 |
+
1000000,
|
| 103 |
+
0.09
|
| 104 |
+
],
|
| 105 |
+
[
|
| 106 |
+
2000000,
|
| 107 |
+
0.04
|
| 108 |
+
],
|
| 109 |
+
[
|
| 110 |
+
4000000,
|
| 111 |
+
0.01
|
| 112 |
+
]
|
| 113 |
+
]
|
| 114 |
+
},
|
| 115 |
+
{
|
| 116 |
+
"name": "wikitext103_low_decay",
|
| 117 |
+
"kind": "anchor_decay",
|
| 118 |
+
"initial": 0.14,
|
| 119 |
+
"final": 0.02,
|
| 120 |
+
"schedule": "log_prefix_anchor",
|
| 121 |
+
"decay_tokens": null,
|
| 122 |
+
"anchors": [
|
| 123 |
+
[
|
| 124 |
+
250000,
|
| 125 |
+
0.14
|
| 126 |
+
],
|
| 127 |
+
[
|
| 128 |
+
500000,
|
| 129 |
+
0.14
|
| 130 |
+
],
|
| 131 |
+
[
|
| 132 |
+
1000000,
|
| 133 |
+
0.1
|
| 134 |
+
],
|
| 135 |
+
[
|
| 136 |
+
2000000,
|
| 137 |
+
0.06
|
| 138 |
+
],
|
| 139 |
+
[
|
| 140 |
+
4000000,
|
| 141 |
+
0.02
|
| 142 |
+
]
|
| 143 |
+
]
|
| 144 |
+
}
|
| 145 |
+
],
|
| 146 |
+
"decay_tokens": null,
|
| 147 |
+
"eval_batches": 64,
|
| 148 |
+
"train_eval_batches": 32,
|
| 149 |
+
"trace_eval_batches": 8,
|
| 150 |
+
"eval_every": 0,
|
| 151 |
+
"log_every": 250,
|
| 152 |
+
"lr": 0.0003,
|
| 153 |
+
"weight_decay": 0.1,
|
| 154 |
+
"grad_clip": 1.0,
|
| 155 |
+
"plateau_delta": 0.01,
|
| 156 |
+
"target_min_dropout": 0.1,
|
| 157 |
+
"min_nonzero_margin": 0.01,
|
| 158 |
+
"min_high_dropout_margin": 0.03,
|
| 159 |
+
"screen_early_stop": false,
|
| 160 |
+
"screen_prune_patience": 3,
|
| 161 |
+
"screen_prune_min_delta": 0.01
|
| 162 |
+
},
|
| 163 |
+
"mode": "locked_stream",
|
| 164 |
+
"seeds": [
|
| 165 |
+
1,
|
| 166 |
+
2,
|
| 167 |
+
3,
|
| 168 |
+
4,
|
| 169 |
+
5
|
| 170 |
+
],
|
| 171 |
+
"models": [
|
| 172 |
+
{
|
| 173 |
+
"model_name": "L12_H8_D320",
|
| 174 |
+
"n_layer": 12,
|
| 175 |
+
"n_head": 8,
|
| 176 |
+
"n_embd": 320
|
| 177 |
+
}
|
| 178 |
+
],
|
| 179 |
+
"device": "mps",
|
| 180 |
+
"torch": "2.12.0",
|
| 181 |
+
"python": "3.11.15 (main, Mar 3 2026, 00:52:57) [Clang 21.0.0 (clang-2100.0.123.102)]",
|
| 182 |
+
"mps_available": true,
|
| 183 |
+
"attribution": "Derived from Andrej Karpathy's nanochat project (https://github.com/karpathy/nanochat), MIT License, Copyright (c) 2025 Andrej Karpathy.",
|
| 184 |
+
"tokenizer_path": ".cache/dropout_decay_wikitext103/tokenizer-v4096.json",
|
| 185 |
+
"encoded_path": ".cache/dropout_decay_wikitext103/tokens-v4096-uint16.npy",
|
| 186 |
+
"train_tokens": 4500020,
|
| 187 |
+
"val_tokens": 500000,
|
| 188 |
+
"effective_token_limits": [
|
| 189 |
+
4500020
|
| 190 |
+
],
|
| 191 |
+
"effective_stream_token_caps": [
|
| 192 |
+
250000,
|
| 193 |
+
500000,
|
| 194 |
+
1000000,
|
| 195 |
+
2000000,
|
| 196 |
+
4000000
|
| 197 |
+
],
|
| 198 |
+
"resume_from": null
|
| 199 |
+
}
|
runs/wikitext103_l12_streaming_validation_5seed/locked_stream/20260531-093525/metrics.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
runs/wikitext103_l12_streaming_validation_5seed/locked_stream/20260531-093525/summary.csv
ADDED
|
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
run_mode,condition,condition_kind,stage,token_limit,model_name,n_layer,n_head,n_embd,parameters,dropout_initial,dropout_final,dropout_schedule,n,mean_train_eval_loss,std_train_eval_loss,mean_val_eval_loss,std_val_eval_loss,mean_generalization_gap,std_generalization_gap
|
| 2 |
+
locked_stream,static_dropout_0,static,0,250000,L12_H8_D320,12,8,320,17367040,0.0,0.0,constant,5,3.5230478435754775,0.03634617773629946,5.340271946787834,0.02699193219713777,1.8172241032123566,0.05178790589031804
|
| 3 |
+
locked_stream,static_dropout_0.02,static,0,250000,L12_H8_D320,12,8,320,17367040,0.02,0.02,constant,5,3.572465108335018,0.027061999609254494,5.274957031011581,0.025475383539799955,1.7024919226765634,0.03656967535619302
|
| 4 |
+
locked_stream,static_dropout_0.04,static,0,250000,L12_H8_D320,12,8,320,17367040,0.04,0.04,constant,5,3.6440532103180887,0.031242522590923742,5.237822580337524,0.018592151453448197,1.593769370019436,0.03528512909650908
|
| 5 |
+
locked_stream,static_dropout_0.06,static,0,250000,L12_H8_D320,12,8,320,17367040,0.06,0.06,constant,5,3.712760145962238,0.018437630763482243,5.215402702987194,0.01734079554543704,1.5026425570249557,0.03154944432034942
|
| 6 |
+
locked_stream,static_dropout_0.08,static,0,250000,L12_H8_D320,12,8,320,17367040,0.08,0.08,constant,5,3.761947725713253,0.017361993699706244,5.189354091882706,0.016114118388750917,1.4274063661694527,0.012218707428925264
|
| 7 |
+
locked_stream,static_dropout_0.1,static,0,250000,L12_H8_D320,12,8,320,17367040,0.1,0.1,constant,5,3.8219083666801454,0.019248328397722874,5.169893845915794,0.023739040366881205,1.3479854792356492,0.03169483270954973
|
| 8 |
+
locked_stream,static_dropout_0.14,static,0,250000,L12_H8_D320,12,8,320,17367040,0.14,0.14,constant,5,3.9050611019134522,0.013283976741307785,5.1635470792651175,0.021953339409384615,1.2584859773516655,0.025343683677925957
|
| 9 |
+
locked_stream,static_dropout_0.18,static,0,250000,L12_H8_D320,12,8,320,17367040,0.18,0.18,constant,5,3.9964362382888794,0.01862825818296394,5.16162400841713,0.014987630048664583,1.16518777012825,0.02013674124988852
|
| 10 |
+
locked_stream,static_dropout_0.2,static,0,250000,L12_H8_D320,12,8,320,17367040,0.2,0.2,constant,5,4.036272630095482,0.013668609797313208,5.170051643252373,0.014071496005973716,1.1337790131568908,0.013715020568700555
|
| 11 |
+
locked_stream,static_dropout_0.26,static,0,250000,L12_H8_D320,12,8,320,17367040,0.26,0.26,constant,5,4.149615630507469,0.024675006962846267,5.194025552272796,0.016128282325428966,1.0444099217653275,0.022196467420881313
|
| 12 |
+
locked_stream,static_dropout_0.3,static,0,250000,L12_H8_D320,12,8,320,17367040,0.3,0.3,constant,5,4.213146212697029,0.013218280901902867,5.214832927286625,0.018060234410376495,1.0016867145895958,0.02723842931339789
|
| 13 |
+
locked_stream,wikitext103_formula_l12,anchor_decay,0,250000,L12_H8_D320,12,8,320,17367040,0.3,0.02,log_prefix_anchor,5,4.213146224617958,0.013218263380840874,5.21483291387558,0.018060268226475495,1.0016866892576217,0.027238437205160136
|
| 14 |
+
locked_stream,wikitext103_low_decay,anchor_decay,0,250000,L12_H8_D320,12,8,320,17367040,0.14,0.02,log_prefix_anchor,5,3.9050611212849615,0.01328398150034886,5.163547059893608,0.02195330299872749,1.2584859386086464,0.025343680837745006
|
| 15 |
+
locked_stream,wikitext103_probe_blend,anchor_decay,0,250000,L12_H8_D320,12,8,320,17367040,0.19,0.01,log_prefix_anchor,5,4.020059056580067,0.016213485710936643,5.165890334546566,0.017075109484847858,1.1458312779664994,0.02016518864116932
|
| 16 |
+
locked_stream,static_dropout_0,static,1,500000,L12_H8_D320,12,8,320,17367040,0.0,0.0,constant,5,3.1505795806646346,0.025699619025332174,5.174065832793713,0.025168511150111032,2.0234862521290777,0.0236535174506497
|
| 17 |
+
locked_stream,static_dropout_0.02,static,1,500000,L12_H8_D320,12,8,320,17367040,0.02,0.02,constant,5,3.261230443418026,0.014168838694048941,5.045114178955555,0.016894569246234076,1.783883735537529,0.019840865900389654
|
| 18 |
+
locked_stream,static_dropout_0.04,static,1,500000,L12_H8_D320,12,8,320,17367040,0.04,0.04,constant,5,3.3515226930379867,0.029537749484816224,4.957330641150475,0.02503345339488141,1.6058079481124878,0.0202359151612378
|
| 19 |
+
locked_stream,static_dropout_0.06,static,1,500000,L12_H8_D320,12,8,320,17367040,0.06,0.06,constant,5,3.4394222095608713,0.013418585935883633,4.894671627879143,0.021331675814859084,1.4552494183182716,0.01946001354297248
|
| 20 |
+
locked_stream,static_dropout_0.08,static,1,500000,L12_H8_D320,12,8,320,17367040,0.08,0.08,constant,5,3.50356438010931,0.024856947856236954,4.857585413753986,0.023903772104159438,1.3540210336446763,0.0310369653543579
|
| 21 |
+
locked_stream,static_dropout_0.1,static,1,500000,L12_H8_D320,12,8,320,17367040,0.1,0.1,constant,5,3.5636927232146265,0.024862277351108786,4.83315060287714,0.025840187319246186,1.2694578796625138,0.01835928668354016
|
| 22 |
+
locked_stream,static_dropout_0.14,static,1,500000,L12_H8_D320,12,8,320,17367040,0.14,0.14,constant,5,3.6649743527173997,0.023120394246437293,4.800087215006352,0.019846558873074828,1.135112862288952,0.012980966896977213
|
| 23 |
+
locked_stream,static_dropout_0.18,static,1,500000,L12_H8_D320,12,8,320,17367040,0.18,0.18,constant,5,3.7571836963295935,0.024732309568709827,4.794638857245445,0.02062163611688947,1.0374551609158515,0.02312552092553772
|
| 24 |
+
locked_stream,static_dropout_0.2,static,1,500000,L12_H8_D320,12,8,320,17367040,0.2,0.2,constant,5,3.7914333969354628,0.018987144953080883,4.787301687896251,0.023601504941398347,0.9958682909607888,0.014572669955232898
|
| 25 |
+
locked_stream,static_dropout_0.26,static,1,500000,L12_H8_D320,12,8,320,17367040,0.26,0.26,constant,5,3.905269515514374,0.02376505939443177,4.8081169500947,0.021576250073037834,0.9028474345803261,0.018700407757224247
|
| 26 |
+
locked_stream,static_dropout_0.3,static,1,500000,L12_H8_D320,12,8,320,17367040,0.3,0.3,constant,5,3.976513123512268,0.010796527385978102,4.82416096329689,0.029627048936626245,0.8476478397846222,0.026893359245285242
|
| 27 |
+
locked_stream,wikitext103_formula_l12,anchor_decay,1,500000,L12_H8_D320,12,8,320,17367040,0.3,0.02,log_prefix_anchor,5,3.918210099637508,0.011987444359375697,4.80534844994545,0.027829930328897787,0.8871383503079414,0.02626272167699895
|
| 28 |
+
locked_stream,wikitext103_low_decay,anchor_decay,1,500000,L12_H8_D320,12,8,320,17367040,0.14,0.02,log_prefix_anchor,5,3.6649743929505347,0.023120354646606448,4.800087235867977,0.01984656043518159,1.1351128429174424,0.01298098541252223
|
| 29 |
+
locked_stream,wikitext103_probe_blend,anchor_decay,1,500000,L12_H8_D320,12,8,320,17367040,0.19,0.01,log_prefix_anchor,5,3.684571087360382,0.02397149156090502,4.7872447416186334,0.0268838649061173,1.1026736542582511,0.021535214146850536
|
| 30 |
+
locked_stream,static_dropout_0,static,2,1000000,L12_H8_D320,12,8,320,17367040,0.0,0.0,constant,5,3.3657630145549775,0.025194332688414405,4.766069588065148,0.03242686059194632,1.40030657351017,0.052237334934698405
|
| 31 |
+
locked_stream,static_dropout_0.02,static,2,1000000,L12_H8_D320,12,8,320,17367040,0.02,0.02,constant,5,3.4323942199349404,0.018124036007212796,4.655830132961273,0.01585338982667534,1.2234359130263328,0.024146051923035036
|
| 32 |
+
locked_stream,static_dropout_0.04,static,2,1000000,L12_H8_D320,12,8,320,17367040,0.04,0.04,constant,5,3.502993068099022,0.017291674978225223,4.595940679311752,0.014619476464690098,1.0929476112127303,0.028916946275823
|
| 33 |
+
locked_stream,static_dropout_0.06,static,2,1000000,L12_H8_D320,12,8,320,17367040,0.06,0.06,constant,5,3.555374290049076,0.018874416206200166,4.557375983893872,0.012636266859277006,1.0020016938447953,0.030341255269329737
|
| 34 |
+
locked_stream,static_dropout_0.08,static,2,1000000,L12_H8_D320,12,8,320,17367040,0.08,0.08,constant,5,3.5999757796525955,0.009078845784580064,4.532606348395348,0.011746366682868653,0.9326305687427521,0.019403758262432406
|
| 35 |
+
locked_stream,static_dropout_0.1,static,2,1000000,L12_H8_D320,12,8,320,17367040,0.1,0.1,constant,5,3.652407945692539,0.022251919818600654,4.518611620366573,0.01351751181040844,0.8662036746740341,0.033766642849086155
|
| 36 |
+
locked_stream,static_dropout_0.14,static,2,1000000,L12_H8_D320,12,8,320,17367040,0.14,0.14,constant,5,3.716250878572464,0.02407671653248301,4.500055414438248,0.014819407090300617,0.7838045358657837,0.038101035301189086
|
| 37 |
+
locked_stream,static_dropout_0.18,static,2,1000000,L12_H8_D320,12,8,320,17367040,0.18,0.18,constant,5,3.786629955470562,0.015769997442310384,4.5023036181926726,0.018490102933264013,0.7156736627221107,0.03180040950921459
|
| 38 |
+
locked_stream,static_dropout_0.2,static,2,1000000,L12_H8_D320,12,8,320,17367040,0.2,0.2,constant,5,3.814779002964497,0.01587114098616348,4.506029562652111,0.02043622151331783,0.6912505596876144,0.03147408021666257
|
| 39 |
+
locked_stream,static_dropout_0.26,static,2,1000000,L12_H8_D320,12,8,320,17367040,0.26,0.26,constant,5,3.9071271896362303,0.02774032700763299,4.52620010226965,0.010149133513914643,0.619072912633419,0.035195638721522884
|
| 40 |
+
locked_stream,static_dropout_0.3,static,2,1000000,L12_H8_D320,12,8,320,17367040,0.3,0.3,constant,5,3.970817744731903,0.017137273586986676,4.546238152682781,0.012672630425565377,0.5754204079508781,0.02420747470801454
|
| 41 |
+
locked_stream,wikitext103_formula_l12,anchor_decay,2,1000000,L12_H8_D320,12,8,320,17367040,0.3,0.02,log_prefix_anchor,5,3.8283171117305757,0.017535628899982475,4.49377838075161,0.014674108647494532,0.6654612690210342,0.02707089996550965
|
| 42 |
+
locked_stream,wikitext103_low_decay,anchor_decay,2,1000000,L12_H8_D320,12,8,320,17367040,0.14,0.02,log_prefix_anchor,5,3.6606912568211554,0.023899445408052746,4.50133255571127,0.01578619433972172,0.8406412988901139,0.03886870429434243
|
| 43 |
+
locked_stream,wikitext103_probe_blend,anchor_decay,2,1000000,L12_H8_D320,12,8,320,17367040,0.19,0.01,log_prefix_anchor,5,3.6495467707514764,0.019490464893964263,4.49399798810482,0.015864095933133645,0.8444512173533439,0.03475725994195922
|
| 44 |
+
locked_stream,static_dropout_0,static,3,2000000,L12_H8_D320,12,8,320,17367040,0.0,0.0,constant,5,3.6409011498093604,0.014124132971310668,4.417352265119552,0.01876483626309062,0.7764511153101921,0.024211298020746064
|
| 45 |
+
locked_stream,static_dropout_0.02,static,3,2000000,L12_H8_D320,12,8,320,17367040,0.02,0.02,constant,5,3.6672953754663467,0.010966930251295475,4.355083760619164,0.026489436313039863,0.6877883851528168,0.026106538952549097
|
| 46 |
+
locked_stream,static_dropout_0.04,static,3,2000000,L12_H8_D320,12,8,320,17367040,0.04,0.04,constant,5,3.703781445324421,0.01970997472996272,4.3262467920780185,0.016533384756725625,0.6224653467535972,0.02577206799431021
|
| 47 |
+
locked_stream,static_dropout_0.06,static,3,2000000,L12_H8_D320,12,8,320,17367040,0.06,0.06,constant,5,3.7317687764763834,0.017157479537739013,4.305900506675243,0.018234543669842573,0.5741317301988602,0.023633182905307998
|
| 48 |
+
locked_stream,static_dropout_0.08,static,3,2000000,L12_H8_D320,12,8,320,17367040,0.08,0.08,constant,5,3.7518652200698854,0.019831550751639896,4.292461846768856,0.016558424281915483,0.5405966266989708,0.034040566494281235
|
| 49 |
+
locked_stream,static_dropout_0.1,static,3,2000000,L12_H8_D320,12,8,320,17367040,0.1,0.1,constant,5,3.7808884128928186,0.01766528157534878,4.285703422129155,0.018994111876547076,0.5048150092363357,0.028072111875115907
|
| 50 |
+
locked_stream,static_dropout_0.14,static,3,2000000,L12_H8_D320,12,8,320,17367040,0.14,0.14,constant,5,3.828673002123833,0.004937581309037719,4.281563547253609,0.02138858846966249,0.452890545129776,0.01755457638984346
|
| 51 |
+
locked_stream,static_dropout_0.18,static,3,2000000,L12_H8_D320,12,8,320,17367040,0.18,0.18,constant,5,3.876796562969685,0.008357602777023191,4.288915157318115,0.019532110178580833,0.4121185943484306,0.013904819825939841
|
| 52 |
+
locked_stream,static_dropout_0.2,static,3,2000000,L12_H8_D320,12,8,320,17367040,0.2,0.2,constant,5,3.898232290148735,0.008893270752025355,4.294206875562668,0.016019909189912813,0.3959745854139328,0.01268490098888804
|
| 53 |
+
locked_stream,static_dropout_0.26,static,3,2000000,L12_H8_D320,12,8,320,17367040,0.26,0.26,constant,5,3.9654731526970863,0.018160558354021094,4.32477528527379,0.017871972287011016,0.359302132576704,0.018728321745075875
|
| 54 |
+
locked_stream,static_dropout_0.3,static,3,2000000,L12_H8_D320,12,8,320,17367040,0.3,0.3,constant,5,4.009651578962803,0.014639843511797184,4.34667577445507,0.015854348211857338,0.3370241954922676,0.019075525784314876
|
| 55 |
+
locked_stream,wikitext103_formula_l12,anchor_decay,3,2000000,L12_H8_D320,12,8,320,17367040,0.3,0.02,log_prefix_anchor,5,3.8088956013321877,0.016376701234484724,4.260725708305836,0.018098739272622868,0.45183010697364806,0.01960955973506382
|
| 56 |
+
locked_stream,wikitext103_low_decay,anchor_decay,3,2000000,L12_H8_D320,12,8,320,17367040,0.14,0.02,log_prefix_anchor,5,3.7474397838115694,0.003830485352807811,4.273568108677864,0.022770462958788148,0.5261283248662949,0.02051193157067845
|
| 57 |
+
locked_stream,wikitext103_probe_blend,anchor_decay,3,2000000,L12_H8_D320,12,8,320,17367040,0.19,0.01,log_prefix_anchor,5,3.720025636255741,0.012889615922098394,4.274319227784872,0.0173783813582558,0.5542935915291309,0.012746502538982045
|
| 58 |
+
locked_stream,static_dropout_0,static,4,4000000,L12_H8_D320,12,8,320,17367040,0.0,0.0,constant,5,3.7750062167644503,0.025453964920340872,4.183486974239349,0.01652781602088682,0.4084807574748993,0.03358628463770181
|
| 59 |
+
locked_stream,static_dropout_0.02,static,4,4000000,L12_H8_D320,12,8,320,17367040,0.02,0.02,constant,5,3.775918734073639,0.03360785862475549,4.145870254188776,0.016455755256661365,0.3699515201151371,0.0368559849747316
|
| 60 |
+
locked_stream,static_dropout_0.04,static,4,4000000,L12_H8_D320,12,8,320,17367040,0.04,0.04,constant,5,3.7978432700037956,0.018504253509089505,4.1331265024840835,0.022679096058077455,0.3352832324802876,0.036956450676459744
|
| 61 |
+
locked_stream,static_dropout_0.06,static,4,4000000,L12_H8_D320,12,8,320,17367040,0.06,0.06,constant,5,3.806561966240406,0.022924307415498075,4.119682380557061,0.008179605355252681,0.3131204143166542,0.028667268842643943
|
| 62 |
+
locked_stream,static_dropout_0.08,static,4,4000000,L12_H8_D320,12,8,320,17367040,0.08,0.08,constant,5,3.8268220275640488,0.021203534366681782,4.111632210761309,0.01857117977689519,0.2848101831972599,0.03679859980086126
|
| 63 |
+
locked_stream,static_dropout_0.1,static,4,4000000,L12_H8_D320,12,8,320,17367040,0.1,0.1,constant,5,3.841722105443478,0.019884552697005082,4.1104619488120075,0.018806640195540816,0.2687398433685303,0.03219305563944893
|
| 64 |
+
locked_stream,static_dropout_0.14,static,4,4000000,L12_H8_D320,12,8,320,17367040,0.14,0.14,constant,5,3.8673695802688597,0.025363134860510252,4.122133788466454,0.015529540565590345,0.2547642081975937,0.0302072467927019
|
| 65 |
+
locked_stream,static_dropout_0.18,static,4,4000000,L12_H8_D320,12,8,320,17367040,0.18,0.18,constant,5,3.9014849200844766,0.018144303418082167,4.130411610752344,0.013005735782293457,0.22892669066786767,0.02966442514236336
|
| 66 |
+
locked_stream,static_dropout_0.2,static,4,4000000,L12_H8_D320,12,8,320,17367040,0.2,0.2,constant,5,3.9154964342713354,0.020116706552102667,4.1393771633505825,0.016742174044810716,0.22388072907924653,0.033233695598367155
|
| 67 |
+
locked_stream,static_dropout_0.26,static,4,4000000,L12_H8_D320,12,8,320,17367040,0.26,0.26,constant,5,3.9775424867868425,0.021977989750098077,4.178351600468159,0.01453281035529064,0.20080911368131638,0.03295020359909027
|
| 68 |
+
locked_stream,static_dropout_0.3,static,4,4000000,L12_H8_D320,12,8,320,17367040,0.3,0.3,constant,5,4.012722708284855,0.024177126067976527,4.194609892368317,0.014131822174274241,0.18188718408346177,0.031100561007876955
|
| 69 |
+
locked_stream,wikitext103_formula_l12,anchor_decay,4,4000000,L12_H8_D320,12,8,320,17367040,0.3,0.02,log_prefix_anchor,5,3.7991417959332465,0.028273745467366926,4.080797865241766,0.019458062276034267,0.2816560693085194,0.038849358858741
|
| 70 |
+
locked_stream,wikitext103_low_decay,anchor_decay,4,4000000,L12_H8_D320,12,8,320,17367040,0.14,0.02,log_prefix_anchor,5,3.776866267621517,0.02647416884246062,4.102013133466244,0.016554390946653275,0.3251468658447266,0.03314100249583806
|
| 71 |
+
locked_stream,wikitext103_probe_blend,anchor_decay,4,4000000,L12_H8_D320,12,8,320,17367040,0.19,0.01,log_prefix_anchor,5,3.7674024373292925,0.027214408565503653,4.096089626103639,0.014523730760550201,0.3286871887743473,0.038218075738052845
|
runs/wikitext103_l12_streaming_validation_5seed/locked_stream/20260531-093525/summary.json
ADDED
|
@@ -0,0 +1,1542 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"run_mode": "locked_stream",
|
| 4 |
+
"condition": "static_dropout_0",
|
| 5 |
+
"condition_kind": "static",
|
| 6 |
+
"stage": 0,
|
| 7 |
+
"token_limit": 250000,
|
| 8 |
+
"model_name": "L12_H8_D320",
|
| 9 |
+
"n_layer": 12,
|
| 10 |
+
"n_head": 8,
|
| 11 |
+
"n_embd": 320,
|
| 12 |
+
"parameters": 17367040,
|
| 13 |
+
"dropout_initial": 0.0,
|
| 14 |
+
"dropout_final": 0.0,
|
| 15 |
+
"dropout_schedule": "constant",
|
| 16 |
+
"n": 5,
|
| 17 |
+
"mean_train_eval_loss": 3.5230478435754775,
|
| 18 |
+
"std_train_eval_loss": 0.03634617773629946,
|
| 19 |
+
"mean_val_eval_loss": 5.340271946787834,
|
| 20 |
+
"std_val_eval_loss": 0.02699193219713777,
|
| 21 |
+
"mean_generalization_gap": 1.8172241032123566,
|
| 22 |
+
"std_generalization_gap": 0.05178790589031804
|
| 23 |
+
},
|
| 24 |
+
{
|
| 25 |
+
"run_mode": "locked_stream",
|
| 26 |
+
"condition": "static_dropout_0.02",
|
| 27 |
+
"condition_kind": "static",
|
| 28 |
+
"stage": 0,
|
| 29 |
+
"token_limit": 250000,
|
| 30 |
+
"model_name": "L12_H8_D320",
|
| 31 |
+
"n_layer": 12,
|
| 32 |
+
"n_head": 8,
|
| 33 |
+
"n_embd": 320,
|
| 34 |
+
"parameters": 17367040,
|
| 35 |
+
"dropout_initial": 0.02,
|
| 36 |
+
"dropout_final": 0.02,
|
| 37 |
+
"dropout_schedule": "constant",
|
| 38 |
+
"n": 5,
|
| 39 |
+
"mean_train_eval_loss": 3.572465108335018,
|
| 40 |
+
"std_train_eval_loss": 0.027061999609254494,
|
| 41 |
+
"mean_val_eval_loss": 5.274957031011581,
|
| 42 |
+
"std_val_eval_loss": 0.025475383539799955,
|
| 43 |
+
"mean_generalization_gap": 1.7024919226765634,
|
| 44 |
+
"std_generalization_gap": 0.03656967535619302
|
| 45 |
+
},
|
| 46 |
+
{
|
| 47 |
+
"run_mode": "locked_stream",
|
| 48 |
+
"condition": "static_dropout_0.04",
|
| 49 |
+
"condition_kind": "static",
|
| 50 |
+
"stage": 0,
|
| 51 |
+
"token_limit": 250000,
|
| 52 |
+
"model_name": "L12_H8_D320",
|
| 53 |
+
"n_layer": 12,
|
| 54 |
+
"n_head": 8,
|
| 55 |
+
"n_embd": 320,
|
| 56 |
+
"parameters": 17367040,
|
| 57 |
+
"dropout_initial": 0.04,
|
| 58 |
+
"dropout_final": 0.04,
|
| 59 |
+
"dropout_schedule": "constant",
|
| 60 |
+
"n": 5,
|
| 61 |
+
"mean_train_eval_loss": 3.6440532103180887,
|
| 62 |
+
"std_train_eval_loss": 0.031242522590923742,
|
| 63 |
+
"mean_val_eval_loss": 5.237822580337524,
|
| 64 |
+
"std_val_eval_loss": 0.018592151453448197,
|
| 65 |
+
"mean_generalization_gap": 1.593769370019436,
|
| 66 |
+
"std_generalization_gap": 0.03528512909650908
|
| 67 |
+
},
|
| 68 |
+
{
|
| 69 |
+
"run_mode": "locked_stream",
|
| 70 |
+
"condition": "static_dropout_0.06",
|
| 71 |
+
"condition_kind": "static",
|
| 72 |
+
"stage": 0,
|
| 73 |
+
"token_limit": 250000,
|
| 74 |
+
"model_name": "L12_H8_D320",
|
| 75 |
+
"n_layer": 12,
|
| 76 |
+
"n_head": 8,
|
| 77 |
+
"n_embd": 320,
|
| 78 |
+
"parameters": 17367040,
|
| 79 |
+
"dropout_initial": 0.06,
|
| 80 |
+
"dropout_final": 0.06,
|
| 81 |
+
"dropout_schedule": "constant",
|
| 82 |
+
"n": 5,
|
| 83 |
+
"mean_train_eval_loss": 3.712760145962238,
|
| 84 |
+
"std_train_eval_loss": 0.018437630763482243,
|
| 85 |
+
"mean_val_eval_loss": 5.215402702987194,
|
| 86 |
+
"std_val_eval_loss": 0.01734079554543704,
|
| 87 |
+
"mean_generalization_gap": 1.5026425570249557,
|
| 88 |
+
"std_generalization_gap": 0.03154944432034942
|
| 89 |
+
},
|
| 90 |
+
{
|
| 91 |
+
"run_mode": "locked_stream",
|
| 92 |
+
"condition": "static_dropout_0.08",
|
| 93 |
+
"condition_kind": "static",
|
| 94 |
+
"stage": 0,
|
| 95 |
+
"token_limit": 250000,
|
| 96 |
+
"model_name": "L12_H8_D320",
|
| 97 |
+
"n_layer": 12,
|
| 98 |
+
"n_head": 8,
|
| 99 |
+
"n_embd": 320,
|
| 100 |
+
"parameters": 17367040,
|
| 101 |
+
"dropout_initial": 0.08,
|
| 102 |
+
"dropout_final": 0.08,
|
| 103 |
+
"dropout_schedule": "constant",
|
| 104 |
+
"n": 5,
|
| 105 |
+
"mean_train_eval_loss": 3.761947725713253,
|
| 106 |
+
"std_train_eval_loss": 0.017361993699706244,
|
| 107 |
+
"mean_val_eval_loss": 5.189354091882706,
|
| 108 |
+
"std_val_eval_loss": 0.016114118388750917,
|
| 109 |
+
"mean_generalization_gap": 1.4274063661694527,
|
| 110 |
+
"std_generalization_gap": 0.012218707428925264
|
| 111 |
+
},
|
| 112 |
+
{
|
| 113 |
+
"run_mode": "locked_stream",
|
| 114 |
+
"condition": "static_dropout_0.1",
|
| 115 |
+
"condition_kind": "static",
|
| 116 |
+
"stage": 0,
|
| 117 |
+
"token_limit": 250000,
|
| 118 |
+
"model_name": "L12_H8_D320",
|
| 119 |
+
"n_layer": 12,
|
| 120 |
+
"n_head": 8,
|
| 121 |
+
"n_embd": 320,
|
| 122 |
+
"parameters": 17367040,
|
| 123 |
+
"dropout_initial": 0.1,
|
| 124 |
+
"dropout_final": 0.1,
|
| 125 |
+
"dropout_schedule": "constant",
|
| 126 |
+
"n": 5,
|
| 127 |
+
"mean_train_eval_loss": 3.8219083666801454,
|
| 128 |
+
"std_train_eval_loss": 0.019248328397722874,
|
| 129 |
+
"mean_val_eval_loss": 5.169893845915794,
|
| 130 |
+
"std_val_eval_loss": 0.023739040366881205,
|
| 131 |
+
"mean_generalization_gap": 1.3479854792356492,
|
| 132 |
+
"std_generalization_gap": 0.03169483270954973
|
| 133 |
+
},
|
| 134 |
+
{
|
| 135 |
+
"run_mode": "locked_stream",
|
| 136 |
+
"condition": "static_dropout_0.14",
|
| 137 |
+
"condition_kind": "static",
|
| 138 |
+
"stage": 0,
|
| 139 |
+
"token_limit": 250000,
|
| 140 |
+
"model_name": "L12_H8_D320",
|
| 141 |
+
"n_layer": 12,
|
| 142 |
+
"n_head": 8,
|
| 143 |
+
"n_embd": 320,
|
| 144 |
+
"parameters": 17367040,
|
| 145 |
+
"dropout_initial": 0.14,
|
| 146 |
+
"dropout_final": 0.14,
|
| 147 |
+
"dropout_schedule": "constant",
|
| 148 |
+
"n": 5,
|
| 149 |
+
"mean_train_eval_loss": 3.9050611019134522,
|
| 150 |
+
"std_train_eval_loss": 0.013283976741307785,
|
| 151 |
+
"mean_val_eval_loss": 5.1635470792651175,
|
| 152 |
+
"std_val_eval_loss": 0.021953339409384615,
|
| 153 |
+
"mean_generalization_gap": 1.2584859773516655,
|
| 154 |
+
"std_generalization_gap": 0.025343683677925957
|
| 155 |
+
},
|
| 156 |
+
{
|
| 157 |
+
"run_mode": "locked_stream",
|
| 158 |
+
"condition": "static_dropout_0.18",
|
| 159 |
+
"condition_kind": "static",
|
| 160 |
+
"stage": 0,
|
| 161 |
+
"token_limit": 250000,
|
| 162 |
+
"model_name": "L12_H8_D320",
|
| 163 |
+
"n_layer": 12,
|
| 164 |
+
"n_head": 8,
|
| 165 |
+
"n_embd": 320,
|
| 166 |
+
"parameters": 17367040,
|
| 167 |
+
"dropout_initial": 0.18,
|
| 168 |
+
"dropout_final": 0.18,
|
| 169 |
+
"dropout_schedule": "constant",
|
| 170 |
+
"n": 5,
|
| 171 |
+
"mean_train_eval_loss": 3.9964362382888794,
|
| 172 |
+
"std_train_eval_loss": 0.01862825818296394,
|
| 173 |
+
"mean_val_eval_loss": 5.16162400841713,
|
| 174 |
+
"std_val_eval_loss": 0.014987630048664583,
|
| 175 |
+
"mean_generalization_gap": 1.16518777012825,
|
| 176 |
+
"std_generalization_gap": 0.02013674124988852
|
| 177 |
+
},
|
| 178 |
+
{
|
| 179 |
+
"run_mode": "locked_stream",
|
| 180 |
+
"condition": "static_dropout_0.2",
|
| 181 |
+
"condition_kind": "static",
|
| 182 |
+
"stage": 0,
|
| 183 |
+
"token_limit": 250000,
|
| 184 |
+
"model_name": "L12_H8_D320",
|
| 185 |
+
"n_layer": 12,
|
| 186 |
+
"n_head": 8,
|
| 187 |
+
"n_embd": 320,
|
| 188 |
+
"parameters": 17367040,
|
| 189 |
+
"dropout_initial": 0.2,
|
| 190 |
+
"dropout_final": 0.2,
|
| 191 |
+
"dropout_schedule": "constant",
|
| 192 |
+
"n": 5,
|
| 193 |
+
"mean_train_eval_loss": 4.036272630095482,
|
| 194 |
+
"std_train_eval_loss": 0.013668609797313208,
|
| 195 |
+
"mean_val_eval_loss": 5.170051643252373,
|
| 196 |
+
"std_val_eval_loss": 0.014071496005973716,
|
| 197 |
+
"mean_generalization_gap": 1.1337790131568908,
|
| 198 |
+
"std_generalization_gap": 0.013715020568700555
|
| 199 |
+
},
|
| 200 |
+
{
|
| 201 |
+
"run_mode": "locked_stream",
|
| 202 |
+
"condition": "static_dropout_0.26",
|
| 203 |
+
"condition_kind": "static",
|
| 204 |
+
"stage": 0,
|
| 205 |
+
"token_limit": 250000,
|
| 206 |
+
"model_name": "L12_H8_D320",
|
| 207 |
+
"n_layer": 12,
|
| 208 |
+
"n_head": 8,
|
| 209 |
+
"n_embd": 320,
|
| 210 |
+
"parameters": 17367040,
|
| 211 |
+
"dropout_initial": 0.26,
|
| 212 |
+
"dropout_final": 0.26,
|
| 213 |
+
"dropout_schedule": "constant",
|
| 214 |
+
"n": 5,
|
| 215 |
+
"mean_train_eval_loss": 4.149615630507469,
|
| 216 |
+
"std_train_eval_loss": 0.024675006962846267,
|
| 217 |
+
"mean_val_eval_loss": 5.194025552272796,
|
| 218 |
+
"std_val_eval_loss": 0.016128282325428966,
|
| 219 |
+
"mean_generalization_gap": 1.0444099217653275,
|
| 220 |
+
"std_generalization_gap": 0.022196467420881313
|
| 221 |
+
},
|
| 222 |
+
{
|
| 223 |
+
"run_mode": "locked_stream",
|
| 224 |
+
"condition": "static_dropout_0.3",
|
| 225 |
+
"condition_kind": "static",
|
| 226 |
+
"stage": 0,
|
| 227 |
+
"token_limit": 250000,
|
| 228 |
+
"model_name": "L12_H8_D320",
|
| 229 |
+
"n_layer": 12,
|
| 230 |
+
"n_head": 8,
|
| 231 |
+
"n_embd": 320,
|
| 232 |
+
"parameters": 17367040,
|
| 233 |
+
"dropout_initial": 0.3,
|
| 234 |
+
"dropout_final": 0.3,
|
| 235 |
+
"dropout_schedule": "constant",
|
| 236 |
+
"n": 5,
|
| 237 |
+
"mean_train_eval_loss": 4.213146212697029,
|
| 238 |
+
"std_train_eval_loss": 0.013218280901902867,
|
| 239 |
+
"mean_val_eval_loss": 5.214832927286625,
|
| 240 |
+
"std_val_eval_loss": 0.018060234410376495,
|
| 241 |
+
"mean_generalization_gap": 1.0016867145895958,
|
| 242 |
+
"std_generalization_gap": 0.02723842931339789
|
| 243 |
+
},
|
| 244 |
+
{
|
| 245 |
+
"run_mode": "locked_stream",
|
| 246 |
+
"condition": "wikitext103_formula_l12",
|
| 247 |
+
"condition_kind": "anchor_decay",
|
| 248 |
+
"stage": 0,
|
| 249 |
+
"token_limit": 250000,
|
| 250 |
+
"model_name": "L12_H8_D320",
|
| 251 |
+
"n_layer": 12,
|
| 252 |
+
"n_head": 8,
|
| 253 |
+
"n_embd": 320,
|
| 254 |
+
"parameters": 17367040,
|
| 255 |
+
"dropout_initial": 0.3,
|
| 256 |
+
"dropout_final": 0.02,
|
| 257 |
+
"dropout_schedule": "log_prefix_anchor",
|
| 258 |
+
"n": 5,
|
| 259 |
+
"mean_train_eval_loss": 4.213146224617958,
|
| 260 |
+
"std_train_eval_loss": 0.013218263380840874,
|
| 261 |
+
"mean_val_eval_loss": 5.21483291387558,
|
| 262 |
+
"std_val_eval_loss": 0.018060268226475495,
|
| 263 |
+
"mean_generalization_gap": 1.0016866892576217,
|
| 264 |
+
"std_generalization_gap": 0.027238437205160136
|
| 265 |
+
},
|
| 266 |
+
{
|
| 267 |
+
"run_mode": "locked_stream",
|
| 268 |
+
"condition": "wikitext103_low_decay",
|
| 269 |
+
"condition_kind": "anchor_decay",
|
| 270 |
+
"stage": 0,
|
| 271 |
+
"token_limit": 250000,
|
| 272 |
+
"model_name": "L12_H8_D320",
|
| 273 |
+
"n_layer": 12,
|
| 274 |
+
"n_head": 8,
|
| 275 |
+
"n_embd": 320,
|
| 276 |
+
"parameters": 17367040,
|
| 277 |
+
"dropout_initial": 0.14,
|
| 278 |
+
"dropout_final": 0.02,
|
| 279 |
+
"dropout_schedule": "log_prefix_anchor",
|
| 280 |
+
"n": 5,
|
| 281 |
+
"mean_train_eval_loss": 3.9050611212849615,
|
| 282 |
+
"std_train_eval_loss": 0.01328398150034886,
|
| 283 |
+
"mean_val_eval_loss": 5.163547059893608,
|
| 284 |
+
"std_val_eval_loss": 0.02195330299872749,
|
| 285 |
+
"mean_generalization_gap": 1.2584859386086464,
|
| 286 |
+
"std_generalization_gap": 0.025343680837745006
|
| 287 |
+
},
|
| 288 |
+
{
|
| 289 |
+
"run_mode": "locked_stream",
|
| 290 |
+
"condition": "wikitext103_probe_blend",
|
| 291 |
+
"condition_kind": "anchor_decay",
|
| 292 |
+
"stage": 0,
|
| 293 |
+
"token_limit": 250000,
|
| 294 |
+
"model_name": "L12_H8_D320",
|
| 295 |
+
"n_layer": 12,
|
| 296 |
+
"n_head": 8,
|
| 297 |
+
"n_embd": 320,
|
| 298 |
+
"parameters": 17367040,
|
| 299 |
+
"dropout_initial": 0.19,
|
| 300 |
+
"dropout_final": 0.01,
|
| 301 |
+
"dropout_schedule": "log_prefix_anchor",
|
| 302 |
+
"n": 5,
|
| 303 |
+
"mean_train_eval_loss": 4.020059056580067,
|
| 304 |
+
"std_train_eval_loss": 0.016213485710936643,
|
| 305 |
+
"mean_val_eval_loss": 5.165890334546566,
|
| 306 |
+
"std_val_eval_loss": 0.017075109484847858,
|
| 307 |
+
"mean_generalization_gap": 1.1458312779664994,
|
| 308 |
+
"std_generalization_gap": 0.02016518864116932
|
| 309 |
+
},
|
| 310 |
+
{
|
| 311 |
+
"run_mode": "locked_stream",
|
| 312 |
+
"condition": "static_dropout_0",
|
| 313 |
+
"condition_kind": "static",
|
| 314 |
+
"stage": 1,
|
| 315 |
+
"token_limit": 500000,
|
| 316 |
+
"model_name": "L12_H8_D320",
|
| 317 |
+
"n_layer": 12,
|
| 318 |
+
"n_head": 8,
|
| 319 |
+
"n_embd": 320,
|
| 320 |
+
"parameters": 17367040,
|
| 321 |
+
"dropout_initial": 0.0,
|
| 322 |
+
"dropout_final": 0.0,
|
| 323 |
+
"dropout_schedule": "constant",
|
| 324 |
+
"n": 5,
|
| 325 |
+
"mean_train_eval_loss": 3.1505795806646346,
|
| 326 |
+
"std_train_eval_loss": 0.025699619025332174,
|
| 327 |
+
"mean_val_eval_loss": 5.174065832793713,
|
| 328 |
+
"std_val_eval_loss": 0.025168511150111032,
|
| 329 |
+
"mean_generalization_gap": 2.0234862521290777,
|
| 330 |
+
"std_generalization_gap": 0.0236535174506497
|
| 331 |
+
},
|
| 332 |
+
{
|
| 333 |
+
"run_mode": "locked_stream",
|
| 334 |
+
"condition": "static_dropout_0.02",
|
| 335 |
+
"condition_kind": "static",
|
| 336 |
+
"stage": 1,
|
| 337 |
+
"token_limit": 500000,
|
| 338 |
+
"model_name": "L12_H8_D320",
|
| 339 |
+
"n_layer": 12,
|
| 340 |
+
"n_head": 8,
|
| 341 |
+
"n_embd": 320,
|
| 342 |
+
"parameters": 17367040,
|
| 343 |
+
"dropout_initial": 0.02,
|
| 344 |
+
"dropout_final": 0.02,
|
| 345 |
+
"dropout_schedule": "constant",
|
| 346 |
+
"n": 5,
|
| 347 |
+
"mean_train_eval_loss": 3.261230443418026,
|
| 348 |
+
"std_train_eval_loss": 0.014168838694048941,
|
| 349 |
+
"mean_val_eval_loss": 5.045114178955555,
|
| 350 |
+
"std_val_eval_loss": 0.016894569246234076,
|
| 351 |
+
"mean_generalization_gap": 1.783883735537529,
|
| 352 |
+
"std_generalization_gap": 0.019840865900389654
|
| 353 |
+
},
|
| 354 |
+
{
|
| 355 |
+
"run_mode": "locked_stream",
|
| 356 |
+
"condition": "static_dropout_0.04",
|
| 357 |
+
"condition_kind": "static",
|
| 358 |
+
"stage": 1,
|
| 359 |
+
"token_limit": 500000,
|
| 360 |
+
"model_name": "L12_H8_D320",
|
| 361 |
+
"n_layer": 12,
|
| 362 |
+
"n_head": 8,
|
| 363 |
+
"n_embd": 320,
|
| 364 |
+
"parameters": 17367040,
|
| 365 |
+
"dropout_initial": 0.04,
|
| 366 |
+
"dropout_final": 0.04,
|
| 367 |
+
"dropout_schedule": "constant",
|
| 368 |
+
"n": 5,
|
| 369 |
+
"mean_train_eval_loss": 3.3515226930379867,
|
| 370 |
+
"std_train_eval_loss": 0.029537749484816224,
|
| 371 |
+
"mean_val_eval_loss": 4.957330641150475,
|
| 372 |
+
"std_val_eval_loss": 0.02503345339488141,
|
| 373 |
+
"mean_generalization_gap": 1.6058079481124878,
|
| 374 |
+
"std_generalization_gap": 0.0202359151612378
|
| 375 |
+
},
|
| 376 |
+
{
|
| 377 |
+
"run_mode": "locked_stream",
|
| 378 |
+
"condition": "static_dropout_0.06",
|
| 379 |
+
"condition_kind": "static",
|
| 380 |
+
"stage": 1,
|
| 381 |
+
"token_limit": 500000,
|
| 382 |
+
"model_name": "L12_H8_D320",
|
| 383 |
+
"n_layer": 12,
|
| 384 |
+
"n_head": 8,
|
| 385 |
+
"n_embd": 320,
|
| 386 |
+
"parameters": 17367040,
|
| 387 |
+
"dropout_initial": 0.06,
|
| 388 |
+
"dropout_final": 0.06,
|
| 389 |
+
"dropout_schedule": "constant",
|
| 390 |
+
"n": 5,
|
| 391 |
+
"mean_train_eval_loss": 3.4394222095608713,
|
| 392 |
+
"std_train_eval_loss": 0.013418585935883633,
|
| 393 |
+
"mean_val_eval_loss": 4.894671627879143,
|
| 394 |
+
"std_val_eval_loss": 0.021331675814859084,
|
| 395 |
+
"mean_generalization_gap": 1.4552494183182716,
|
| 396 |
+
"std_generalization_gap": 0.01946001354297248
|
| 397 |
+
},
|
| 398 |
+
{
|
| 399 |
+
"run_mode": "locked_stream",
|
| 400 |
+
"condition": "static_dropout_0.08",
|
| 401 |
+
"condition_kind": "static",
|
| 402 |
+
"stage": 1,
|
| 403 |
+
"token_limit": 500000,
|
| 404 |
+
"model_name": "L12_H8_D320",
|
| 405 |
+
"n_layer": 12,
|
| 406 |
+
"n_head": 8,
|
| 407 |
+
"n_embd": 320,
|
| 408 |
+
"parameters": 17367040,
|
| 409 |
+
"dropout_initial": 0.08,
|
| 410 |
+
"dropout_final": 0.08,
|
| 411 |
+
"dropout_schedule": "constant",
|
| 412 |
+
"n": 5,
|
| 413 |
+
"mean_train_eval_loss": 3.50356438010931,
|
| 414 |
+
"std_train_eval_loss": 0.024856947856236954,
|
| 415 |
+
"mean_val_eval_loss": 4.857585413753986,
|
| 416 |
+
"std_val_eval_loss": 0.023903772104159438,
|
| 417 |
+
"mean_generalization_gap": 1.3540210336446763,
|
| 418 |
+
"std_generalization_gap": 0.0310369653543579
|
| 419 |
+
},
|
| 420 |
+
{
|
| 421 |
+
"run_mode": "locked_stream",
|
| 422 |
+
"condition": "static_dropout_0.1",
|
| 423 |
+
"condition_kind": "static",
|
| 424 |
+
"stage": 1,
|
| 425 |
+
"token_limit": 500000,
|
| 426 |
+
"model_name": "L12_H8_D320",
|
| 427 |
+
"n_layer": 12,
|
| 428 |
+
"n_head": 8,
|
| 429 |
+
"n_embd": 320,
|
| 430 |
+
"parameters": 17367040,
|
| 431 |
+
"dropout_initial": 0.1,
|
| 432 |
+
"dropout_final": 0.1,
|
| 433 |
+
"dropout_schedule": "constant",
|
| 434 |
+
"n": 5,
|
| 435 |
+
"mean_train_eval_loss": 3.5636927232146265,
|
| 436 |
+
"std_train_eval_loss": 0.024862277351108786,
|
| 437 |
+
"mean_val_eval_loss": 4.83315060287714,
|
| 438 |
+
"std_val_eval_loss": 0.025840187319246186,
|
| 439 |
+
"mean_generalization_gap": 1.2694578796625138,
|
| 440 |
+
"std_generalization_gap": 0.01835928668354016
|
| 441 |
+
},
|
| 442 |
+
{
|
| 443 |
+
"run_mode": "locked_stream",
|
| 444 |
+
"condition": "static_dropout_0.14",
|
| 445 |
+
"condition_kind": "static",
|
| 446 |
+
"stage": 1,
|
| 447 |
+
"token_limit": 500000,
|
| 448 |
+
"model_name": "L12_H8_D320",
|
| 449 |
+
"n_layer": 12,
|
| 450 |
+
"n_head": 8,
|
| 451 |
+
"n_embd": 320,
|
| 452 |
+
"parameters": 17367040,
|
| 453 |
+
"dropout_initial": 0.14,
|
| 454 |
+
"dropout_final": 0.14,
|
| 455 |
+
"dropout_schedule": "constant",
|
| 456 |
+
"n": 5,
|
| 457 |
+
"mean_train_eval_loss": 3.6649743527173997,
|
| 458 |
+
"std_train_eval_loss": 0.023120394246437293,
|
| 459 |
+
"mean_val_eval_loss": 4.800087215006352,
|
| 460 |
+
"std_val_eval_loss": 0.019846558873074828,
|
| 461 |
+
"mean_generalization_gap": 1.135112862288952,
|
| 462 |
+
"std_generalization_gap": 0.012980966896977213
|
| 463 |
+
},
|
| 464 |
+
{
|
| 465 |
+
"run_mode": "locked_stream",
|
| 466 |
+
"condition": "static_dropout_0.18",
|
| 467 |
+
"condition_kind": "static",
|
| 468 |
+
"stage": 1,
|
| 469 |
+
"token_limit": 500000,
|
| 470 |
+
"model_name": "L12_H8_D320",
|
| 471 |
+
"n_layer": 12,
|
| 472 |
+
"n_head": 8,
|
| 473 |
+
"n_embd": 320,
|
| 474 |
+
"parameters": 17367040,
|
| 475 |
+
"dropout_initial": 0.18,
|
| 476 |
+
"dropout_final": 0.18,
|
| 477 |
+
"dropout_schedule": "constant",
|
| 478 |
+
"n": 5,
|
| 479 |
+
"mean_train_eval_loss": 3.7571836963295935,
|
| 480 |
+
"std_train_eval_loss": 0.024732309568709827,
|
| 481 |
+
"mean_val_eval_loss": 4.794638857245445,
|
| 482 |
+
"std_val_eval_loss": 0.02062163611688947,
|
| 483 |
+
"mean_generalization_gap": 1.0374551609158515,
|
| 484 |
+
"std_generalization_gap": 0.02312552092553772
|
| 485 |
+
},
|
| 486 |
+
{
|
| 487 |
+
"run_mode": "locked_stream",
|
| 488 |
+
"condition": "static_dropout_0.2",
|
| 489 |
+
"condition_kind": "static",
|
| 490 |
+
"stage": 1,
|
| 491 |
+
"token_limit": 500000,
|
| 492 |
+
"model_name": "L12_H8_D320",
|
| 493 |
+
"n_layer": 12,
|
| 494 |
+
"n_head": 8,
|
| 495 |
+
"n_embd": 320,
|
| 496 |
+
"parameters": 17367040,
|
| 497 |
+
"dropout_initial": 0.2,
|
| 498 |
+
"dropout_final": 0.2,
|
| 499 |
+
"dropout_schedule": "constant",
|
| 500 |
+
"n": 5,
|
| 501 |
+
"mean_train_eval_loss": 3.7914333969354628,
|
| 502 |
+
"std_train_eval_loss": 0.018987144953080883,
|
| 503 |
+
"mean_val_eval_loss": 4.787301687896251,
|
| 504 |
+
"std_val_eval_loss": 0.023601504941398347,
|
| 505 |
+
"mean_generalization_gap": 0.9958682909607888,
|
| 506 |
+
"std_generalization_gap": 0.014572669955232898
|
| 507 |
+
},
|
| 508 |
+
{
|
| 509 |
+
"run_mode": "locked_stream",
|
| 510 |
+
"condition": "static_dropout_0.26",
|
| 511 |
+
"condition_kind": "static",
|
| 512 |
+
"stage": 1,
|
| 513 |
+
"token_limit": 500000,
|
| 514 |
+
"model_name": "L12_H8_D320",
|
| 515 |
+
"n_layer": 12,
|
| 516 |
+
"n_head": 8,
|
| 517 |
+
"n_embd": 320,
|
| 518 |
+
"parameters": 17367040,
|
| 519 |
+
"dropout_initial": 0.26,
|
| 520 |
+
"dropout_final": 0.26,
|
| 521 |
+
"dropout_schedule": "constant",
|
| 522 |
+
"n": 5,
|
| 523 |
+
"mean_train_eval_loss": 3.905269515514374,
|
| 524 |
+
"std_train_eval_loss": 0.02376505939443177,
|
| 525 |
+
"mean_val_eval_loss": 4.8081169500947,
|
| 526 |
+
"std_val_eval_loss": 0.021576250073037834,
|
| 527 |
+
"mean_generalization_gap": 0.9028474345803261,
|
| 528 |
+
"std_generalization_gap": 0.018700407757224247
|
| 529 |
+
},
|
| 530 |
+
{
|
| 531 |
+
"run_mode": "locked_stream",
|
| 532 |
+
"condition": "static_dropout_0.3",
|
| 533 |
+
"condition_kind": "static",
|
| 534 |
+
"stage": 1,
|
| 535 |
+
"token_limit": 500000,
|
| 536 |
+
"model_name": "L12_H8_D320",
|
| 537 |
+
"n_layer": 12,
|
| 538 |
+
"n_head": 8,
|
| 539 |
+
"n_embd": 320,
|
| 540 |
+
"parameters": 17367040,
|
| 541 |
+
"dropout_initial": 0.3,
|
| 542 |
+
"dropout_final": 0.3,
|
| 543 |
+
"dropout_schedule": "constant",
|
| 544 |
+
"n": 5,
|
| 545 |
+
"mean_train_eval_loss": 3.976513123512268,
|
| 546 |
+
"std_train_eval_loss": 0.010796527385978102,
|
| 547 |
+
"mean_val_eval_loss": 4.82416096329689,
|
| 548 |
+
"std_val_eval_loss": 0.029627048936626245,
|
| 549 |
+
"mean_generalization_gap": 0.8476478397846222,
|
| 550 |
+
"std_generalization_gap": 0.026893359245285242
|
| 551 |
+
},
|
| 552 |
+
{
|
| 553 |
+
"run_mode": "locked_stream",
|
| 554 |
+
"condition": "wikitext103_formula_l12",
|
| 555 |
+
"condition_kind": "anchor_decay",
|
| 556 |
+
"stage": 1,
|
| 557 |
+
"token_limit": 500000,
|
| 558 |
+
"model_name": "L12_H8_D320",
|
| 559 |
+
"n_layer": 12,
|
| 560 |
+
"n_head": 8,
|
| 561 |
+
"n_embd": 320,
|
| 562 |
+
"parameters": 17367040,
|
| 563 |
+
"dropout_initial": 0.3,
|
| 564 |
+
"dropout_final": 0.02,
|
| 565 |
+
"dropout_schedule": "log_prefix_anchor",
|
| 566 |
+
"n": 5,
|
| 567 |
+
"mean_train_eval_loss": 3.918210099637508,
|
| 568 |
+
"std_train_eval_loss": 0.011987444359375697,
|
| 569 |
+
"mean_val_eval_loss": 4.80534844994545,
|
| 570 |
+
"std_val_eval_loss": 0.027829930328897787,
|
| 571 |
+
"mean_generalization_gap": 0.8871383503079414,
|
| 572 |
+
"std_generalization_gap": 0.02626272167699895
|
| 573 |
+
},
|
| 574 |
+
{
|
| 575 |
+
"run_mode": "locked_stream",
|
| 576 |
+
"condition": "wikitext103_low_decay",
|
| 577 |
+
"condition_kind": "anchor_decay",
|
| 578 |
+
"stage": 1,
|
| 579 |
+
"token_limit": 500000,
|
| 580 |
+
"model_name": "L12_H8_D320",
|
| 581 |
+
"n_layer": 12,
|
| 582 |
+
"n_head": 8,
|
| 583 |
+
"n_embd": 320,
|
| 584 |
+
"parameters": 17367040,
|
| 585 |
+
"dropout_initial": 0.14,
|
| 586 |
+
"dropout_final": 0.02,
|
| 587 |
+
"dropout_schedule": "log_prefix_anchor",
|
| 588 |
+
"n": 5,
|
| 589 |
+
"mean_train_eval_loss": 3.6649743929505347,
|
| 590 |
+
"std_train_eval_loss": 0.023120354646606448,
|
| 591 |
+
"mean_val_eval_loss": 4.800087235867977,
|
| 592 |
+
"std_val_eval_loss": 0.01984656043518159,
|
| 593 |
+
"mean_generalization_gap": 1.1351128429174424,
|
| 594 |
+
"std_generalization_gap": 0.01298098541252223
|
| 595 |
+
},
|
| 596 |
+
{
|
| 597 |
+
"run_mode": "locked_stream",
|
| 598 |
+
"condition": "wikitext103_probe_blend",
|
| 599 |
+
"condition_kind": "anchor_decay",
|
| 600 |
+
"stage": 1,
|
| 601 |
+
"token_limit": 500000,
|
| 602 |
+
"model_name": "L12_H8_D320",
|
| 603 |
+
"n_layer": 12,
|
| 604 |
+
"n_head": 8,
|
| 605 |
+
"n_embd": 320,
|
| 606 |
+
"parameters": 17367040,
|
| 607 |
+
"dropout_initial": 0.19,
|
| 608 |
+
"dropout_final": 0.01,
|
| 609 |
+
"dropout_schedule": "log_prefix_anchor",
|
| 610 |
+
"n": 5,
|
| 611 |
+
"mean_train_eval_loss": 3.684571087360382,
|
| 612 |
+
"std_train_eval_loss": 0.02397149156090502,
|
| 613 |
+
"mean_val_eval_loss": 4.7872447416186334,
|
| 614 |
+
"std_val_eval_loss": 0.0268838649061173,
|
| 615 |
+
"mean_generalization_gap": 1.1026736542582511,
|
| 616 |
+
"std_generalization_gap": 0.021535214146850536
|
| 617 |
+
},
|
| 618 |
+
{
|
| 619 |
+
"run_mode": "locked_stream",
|
| 620 |
+
"condition": "static_dropout_0",
|
| 621 |
+
"condition_kind": "static",
|
| 622 |
+
"stage": 2,
|
| 623 |
+
"token_limit": 1000000,
|
| 624 |
+
"model_name": "L12_H8_D320",
|
| 625 |
+
"n_layer": 12,
|
| 626 |
+
"n_head": 8,
|
| 627 |
+
"n_embd": 320,
|
| 628 |
+
"parameters": 17367040,
|
| 629 |
+
"dropout_initial": 0.0,
|
| 630 |
+
"dropout_final": 0.0,
|
| 631 |
+
"dropout_schedule": "constant",
|
| 632 |
+
"n": 5,
|
| 633 |
+
"mean_train_eval_loss": 3.3657630145549775,
|
| 634 |
+
"std_train_eval_loss": 0.025194332688414405,
|
| 635 |
+
"mean_val_eval_loss": 4.766069588065148,
|
| 636 |
+
"std_val_eval_loss": 0.03242686059194632,
|
| 637 |
+
"mean_generalization_gap": 1.40030657351017,
|
| 638 |
+
"std_generalization_gap": 0.052237334934698405
|
| 639 |
+
},
|
| 640 |
+
{
|
| 641 |
+
"run_mode": "locked_stream",
|
| 642 |
+
"condition": "static_dropout_0.02",
|
| 643 |
+
"condition_kind": "static",
|
| 644 |
+
"stage": 2,
|
| 645 |
+
"token_limit": 1000000,
|
| 646 |
+
"model_name": "L12_H8_D320",
|
| 647 |
+
"n_layer": 12,
|
| 648 |
+
"n_head": 8,
|
| 649 |
+
"n_embd": 320,
|
| 650 |
+
"parameters": 17367040,
|
| 651 |
+
"dropout_initial": 0.02,
|
| 652 |
+
"dropout_final": 0.02,
|
| 653 |
+
"dropout_schedule": "constant",
|
| 654 |
+
"n": 5,
|
| 655 |
+
"mean_train_eval_loss": 3.4323942199349404,
|
| 656 |
+
"std_train_eval_loss": 0.018124036007212796,
|
| 657 |
+
"mean_val_eval_loss": 4.655830132961273,
|
| 658 |
+
"std_val_eval_loss": 0.01585338982667534,
|
| 659 |
+
"mean_generalization_gap": 1.2234359130263328,
|
| 660 |
+
"std_generalization_gap": 0.024146051923035036
|
| 661 |
+
},
|
| 662 |
+
{
|
| 663 |
+
"run_mode": "locked_stream",
|
| 664 |
+
"condition": "static_dropout_0.04",
|
| 665 |
+
"condition_kind": "static",
|
| 666 |
+
"stage": 2,
|
| 667 |
+
"token_limit": 1000000,
|
| 668 |
+
"model_name": "L12_H8_D320",
|
| 669 |
+
"n_layer": 12,
|
| 670 |
+
"n_head": 8,
|
| 671 |
+
"n_embd": 320,
|
| 672 |
+
"parameters": 17367040,
|
| 673 |
+
"dropout_initial": 0.04,
|
| 674 |
+
"dropout_final": 0.04,
|
| 675 |
+
"dropout_schedule": "constant",
|
| 676 |
+
"n": 5,
|
| 677 |
+
"mean_train_eval_loss": 3.502993068099022,
|
| 678 |
+
"std_train_eval_loss": 0.017291674978225223,
|
| 679 |
+
"mean_val_eval_loss": 4.595940679311752,
|
| 680 |
+
"std_val_eval_loss": 0.014619476464690098,
|
| 681 |
+
"mean_generalization_gap": 1.0929476112127303,
|
| 682 |
+
"std_generalization_gap": 0.028916946275823
|
| 683 |
+
},
|
| 684 |
+
{
|
| 685 |
+
"run_mode": "locked_stream",
|
| 686 |
+
"condition": "static_dropout_0.06",
|
| 687 |
+
"condition_kind": "static",
|
| 688 |
+
"stage": 2,
|
| 689 |
+
"token_limit": 1000000,
|
| 690 |
+
"model_name": "L12_H8_D320",
|
| 691 |
+
"n_layer": 12,
|
| 692 |
+
"n_head": 8,
|
| 693 |
+
"n_embd": 320,
|
| 694 |
+
"parameters": 17367040,
|
| 695 |
+
"dropout_initial": 0.06,
|
| 696 |
+
"dropout_final": 0.06,
|
| 697 |
+
"dropout_schedule": "constant",
|
| 698 |
+
"n": 5,
|
| 699 |
+
"mean_train_eval_loss": 3.555374290049076,
|
| 700 |
+
"std_train_eval_loss": 0.018874416206200166,
|
| 701 |
+
"mean_val_eval_loss": 4.557375983893872,
|
| 702 |
+
"std_val_eval_loss": 0.012636266859277006,
|
| 703 |
+
"mean_generalization_gap": 1.0020016938447953,
|
| 704 |
+
"std_generalization_gap": 0.030341255269329737
|
| 705 |
+
},
|
| 706 |
+
{
|
| 707 |
+
"run_mode": "locked_stream",
|
| 708 |
+
"condition": "static_dropout_0.08",
|
| 709 |
+
"condition_kind": "static",
|
| 710 |
+
"stage": 2,
|
| 711 |
+
"token_limit": 1000000,
|
| 712 |
+
"model_name": "L12_H8_D320",
|
| 713 |
+
"n_layer": 12,
|
| 714 |
+
"n_head": 8,
|
| 715 |
+
"n_embd": 320,
|
| 716 |
+
"parameters": 17367040,
|
| 717 |
+
"dropout_initial": 0.08,
|
| 718 |
+
"dropout_final": 0.08,
|
| 719 |
+
"dropout_schedule": "constant",
|
| 720 |
+
"n": 5,
|
| 721 |
+
"mean_train_eval_loss": 3.5999757796525955,
|
| 722 |
+
"std_train_eval_loss": 0.009078845784580064,
|
| 723 |
+
"mean_val_eval_loss": 4.532606348395348,
|
| 724 |
+
"std_val_eval_loss": 0.011746366682868653,
|
| 725 |
+
"mean_generalization_gap": 0.9326305687427521,
|
| 726 |
+
"std_generalization_gap": 0.019403758262432406
|
| 727 |
+
},
|
| 728 |
+
{
|
| 729 |
+
"run_mode": "locked_stream",
|
| 730 |
+
"condition": "static_dropout_0.1",
|
| 731 |
+
"condition_kind": "static",
|
| 732 |
+
"stage": 2,
|
| 733 |
+
"token_limit": 1000000,
|
| 734 |
+
"model_name": "L12_H8_D320",
|
| 735 |
+
"n_layer": 12,
|
| 736 |
+
"n_head": 8,
|
| 737 |
+
"n_embd": 320,
|
| 738 |
+
"parameters": 17367040,
|
| 739 |
+
"dropout_initial": 0.1,
|
| 740 |
+
"dropout_final": 0.1,
|
| 741 |
+
"dropout_schedule": "constant",
|
| 742 |
+
"n": 5,
|
| 743 |
+
"mean_train_eval_loss": 3.652407945692539,
|
| 744 |
+
"std_train_eval_loss": 0.022251919818600654,
|
| 745 |
+
"mean_val_eval_loss": 4.518611620366573,
|
| 746 |
+
"std_val_eval_loss": 0.01351751181040844,
|
| 747 |
+
"mean_generalization_gap": 0.8662036746740341,
|
| 748 |
+
"std_generalization_gap": 0.033766642849086155
|
| 749 |
+
},
|
| 750 |
+
{
|
| 751 |
+
"run_mode": "locked_stream",
|
| 752 |
+
"condition": "static_dropout_0.14",
|
| 753 |
+
"condition_kind": "static",
|
| 754 |
+
"stage": 2,
|
| 755 |
+
"token_limit": 1000000,
|
| 756 |
+
"model_name": "L12_H8_D320",
|
| 757 |
+
"n_layer": 12,
|
| 758 |
+
"n_head": 8,
|
| 759 |
+
"n_embd": 320,
|
| 760 |
+
"parameters": 17367040,
|
| 761 |
+
"dropout_initial": 0.14,
|
| 762 |
+
"dropout_final": 0.14,
|
| 763 |
+
"dropout_schedule": "constant",
|
| 764 |
+
"n": 5,
|
| 765 |
+
"mean_train_eval_loss": 3.716250878572464,
|
| 766 |
+
"std_train_eval_loss": 0.02407671653248301,
|
| 767 |
+
"mean_val_eval_loss": 4.500055414438248,
|
| 768 |
+
"std_val_eval_loss": 0.014819407090300617,
|
| 769 |
+
"mean_generalization_gap": 0.7838045358657837,
|
| 770 |
+
"std_generalization_gap": 0.038101035301189086
|
| 771 |
+
},
|
| 772 |
+
{
|
| 773 |
+
"run_mode": "locked_stream",
|
| 774 |
+
"condition": "static_dropout_0.18",
|
| 775 |
+
"condition_kind": "static",
|
| 776 |
+
"stage": 2,
|
| 777 |
+
"token_limit": 1000000,
|
| 778 |
+
"model_name": "L12_H8_D320",
|
| 779 |
+
"n_layer": 12,
|
| 780 |
+
"n_head": 8,
|
| 781 |
+
"n_embd": 320,
|
| 782 |
+
"parameters": 17367040,
|
| 783 |
+
"dropout_initial": 0.18,
|
| 784 |
+
"dropout_final": 0.18,
|
| 785 |
+
"dropout_schedule": "constant",
|
| 786 |
+
"n": 5,
|
| 787 |
+
"mean_train_eval_loss": 3.786629955470562,
|
| 788 |
+
"std_train_eval_loss": 0.015769997442310384,
|
| 789 |
+
"mean_val_eval_loss": 4.5023036181926726,
|
| 790 |
+
"std_val_eval_loss": 0.018490102933264013,
|
| 791 |
+
"mean_generalization_gap": 0.7156736627221107,
|
| 792 |
+
"std_generalization_gap": 0.03180040950921459
|
| 793 |
+
},
|
| 794 |
+
{
|
| 795 |
+
"run_mode": "locked_stream",
|
| 796 |
+
"condition": "static_dropout_0.2",
|
| 797 |
+
"condition_kind": "static",
|
| 798 |
+
"stage": 2,
|
| 799 |
+
"token_limit": 1000000,
|
| 800 |
+
"model_name": "L12_H8_D320",
|
| 801 |
+
"n_layer": 12,
|
| 802 |
+
"n_head": 8,
|
| 803 |
+
"n_embd": 320,
|
| 804 |
+
"parameters": 17367040,
|
| 805 |
+
"dropout_initial": 0.2,
|
| 806 |
+
"dropout_final": 0.2,
|
| 807 |
+
"dropout_schedule": "constant",
|
| 808 |
+
"n": 5,
|
| 809 |
+
"mean_train_eval_loss": 3.814779002964497,
|
| 810 |
+
"std_train_eval_loss": 0.01587114098616348,
|
| 811 |
+
"mean_val_eval_loss": 4.506029562652111,
|
| 812 |
+
"std_val_eval_loss": 0.02043622151331783,
|
| 813 |
+
"mean_generalization_gap": 0.6912505596876144,
|
| 814 |
+
"std_generalization_gap": 0.03147408021666257
|
| 815 |
+
},
|
| 816 |
+
{
|
| 817 |
+
"run_mode": "locked_stream",
|
| 818 |
+
"condition": "static_dropout_0.26",
|
| 819 |
+
"condition_kind": "static",
|
| 820 |
+
"stage": 2,
|
| 821 |
+
"token_limit": 1000000,
|
| 822 |
+
"model_name": "L12_H8_D320",
|
| 823 |
+
"n_layer": 12,
|
| 824 |
+
"n_head": 8,
|
| 825 |
+
"n_embd": 320,
|
| 826 |
+
"parameters": 17367040,
|
| 827 |
+
"dropout_initial": 0.26,
|
| 828 |
+
"dropout_final": 0.26,
|
| 829 |
+
"dropout_schedule": "constant",
|
| 830 |
+
"n": 5,
|
| 831 |
+
"mean_train_eval_loss": 3.9071271896362303,
|
| 832 |
+
"std_train_eval_loss": 0.02774032700763299,
|
| 833 |
+
"mean_val_eval_loss": 4.52620010226965,
|
| 834 |
+
"std_val_eval_loss": 0.010149133513914643,
|
| 835 |
+
"mean_generalization_gap": 0.619072912633419,
|
| 836 |
+
"std_generalization_gap": 0.035195638721522884
|
| 837 |
+
},
|
| 838 |
+
{
|
| 839 |
+
"run_mode": "locked_stream",
|
| 840 |
+
"condition": "static_dropout_0.3",
|
| 841 |
+
"condition_kind": "static",
|
| 842 |
+
"stage": 2,
|
| 843 |
+
"token_limit": 1000000,
|
| 844 |
+
"model_name": "L12_H8_D320",
|
| 845 |
+
"n_layer": 12,
|
| 846 |
+
"n_head": 8,
|
| 847 |
+
"n_embd": 320,
|
| 848 |
+
"parameters": 17367040,
|
| 849 |
+
"dropout_initial": 0.3,
|
| 850 |
+
"dropout_final": 0.3,
|
| 851 |
+
"dropout_schedule": "constant",
|
| 852 |
+
"n": 5,
|
| 853 |
+
"mean_train_eval_loss": 3.970817744731903,
|
| 854 |
+
"std_train_eval_loss": 0.017137273586986676,
|
| 855 |
+
"mean_val_eval_loss": 4.546238152682781,
|
| 856 |
+
"std_val_eval_loss": 0.012672630425565377,
|
| 857 |
+
"mean_generalization_gap": 0.5754204079508781,
|
| 858 |
+
"std_generalization_gap": 0.02420747470801454
|
| 859 |
+
},
|
| 860 |
+
{
|
| 861 |
+
"run_mode": "locked_stream",
|
| 862 |
+
"condition": "wikitext103_formula_l12",
|
| 863 |
+
"condition_kind": "anchor_decay",
|
| 864 |
+
"stage": 2,
|
| 865 |
+
"token_limit": 1000000,
|
| 866 |
+
"model_name": "L12_H8_D320",
|
| 867 |
+
"n_layer": 12,
|
| 868 |
+
"n_head": 8,
|
| 869 |
+
"n_embd": 320,
|
| 870 |
+
"parameters": 17367040,
|
| 871 |
+
"dropout_initial": 0.3,
|
| 872 |
+
"dropout_final": 0.02,
|
| 873 |
+
"dropout_schedule": "log_prefix_anchor",
|
| 874 |
+
"n": 5,
|
| 875 |
+
"mean_train_eval_loss": 3.8283171117305757,
|
| 876 |
+
"std_train_eval_loss": 0.017535628899982475,
|
| 877 |
+
"mean_val_eval_loss": 4.49377838075161,
|
| 878 |
+
"std_val_eval_loss": 0.014674108647494532,
|
| 879 |
+
"mean_generalization_gap": 0.6654612690210342,
|
| 880 |
+
"std_generalization_gap": 0.02707089996550965
|
| 881 |
+
},
|
| 882 |
+
{
|
| 883 |
+
"run_mode": "locked_stream",
|
| 884 |
+
"condition": "wikitext103_low_decay",
|
| 885 |
+
"condition_kind": "anchor_decay",
|
| 886 |
+
"stage": 2,
|
| 887 |
+
"token_limit": 1000000,
|
| 888 |
+
"model_name": "L12_H8_D320",
|
| 889 |
+
"n_layer": 12,
|
| 890 |
+
"n_head": 8,
|
| 891 |
+
"n_embd": 320,
|
| 892 |
+
"parameters": 17367040,
|
| 893 |
+
"dropout_initial": 0.14,
|
| 894 |
+
"dropout_final": 0.02,
|
| 895 |
+
"dropout_schedule": "log_prefix_anchor",
|
| 896 |
+
"n": 5,
|
| 897 |
+
"mean_train_eval_loss": 3.6606912568211554,
|
| 898 |
+
"std_train_eval_loss": 0.023899445408052746,
|
| 899 |
+
"mean_val_eval_loss": 4.50133255571127,
|
| 900 |
+
"std_val_eval_loss": 0.01578619433972172,
|
| 901 |
+
"mean_generalization_gap": 0.8406412988901139,
|
| 902 |
+
"std_generalization_gap": 0.03886870429434243
|
| 903 |
+
},
|
| 904 |
+
{
|
| 905 |
+
"run_mode": "locked_stream",
|
| 906 |
+
"condition": "wikitext103_probe_blend",
|
| 907 |
+
"condition_kind": "anchor_decay",
|
| 908 |
+
"stage": 2,
|
| 909 |
+
"token_limit": 1000000,
|
| 910 |
+
"model_name": "L12_H8_D320",
|
| 911 |
+
"n_layer": 12,
|
| 912 |
+
"n_head": 8,
|
| 913 |
+
"n_embd": 320,
|
| 914 |
+
"parameters": 17367040,
|
| 915 |
+
"dropout_initial": 0.19,
|
| 916 |
+
"dropout_final": 0.01,
|
| 917 |
+
"dropout_schedule": "log_prefix_anchor",
|
| 918 |
+
"n": 5,
|
| 919 |
+
"mean_train_eval_loss": 3.6495467707514764,
|
| 920 |
+
"std_train_eval_loss": 0.019490464893964263,
|
| 921 |
+
"mean_val_eval_loss": 4.49399798810482,
|
| 922 |
+
"std_val_eval_loss": 0.015864095933133645,
|
| 923 |
+
"mean_generalization_gap": 0.8444512173533439,
|
| 924 |
+
"std_generalization_gap": 0.03475725994195922
|
| 925 |
+
},
|
| 926 |
+
{
|
| 927 |
+
"run_mode": "locked_stream",
|
| 928 |
+
"condition": "static_dropout_0",
|
| 929 |
+
"condition_kind": "static",
|
| 930 |
+
"stage": 3,
|
| 931 |
+
"token_limit": 2000000,
|
| 932 |
+
"model_name": "L12_H8_D320",
|
| 933 |
+
"n_layer": 12,
|
| 934 |
+
"n_head": 8,
|
| 935 |
+
"n_embd": 320,
|
| 936 |
+
"parameters": 17367040,
|
| 937 |
+
"dropout_initial": 0.0,
|
| 938 |
+
"dropout_final": 0.0,
|
| 939 |
+
"dropout_schedule": "constant",
|
| 940 |
+
"n": 5,
|
| 941 |
+
"mean_train_eval_loss": 3.6409011498093604,
|
| 942 |
+
"std_train_eval_loss": 0.014124132971310668,
|
| 943 |
+
"mean_val_eval_loss": 4.417352265119552,
|
| 944 |
+
"std_val_eval_loss": 0.01876483626309062,
|
| 945 |
+
"mean_generalization_gap": 0.7764511153101921,
|
| 946 |
+
"std_generalization_gap": 0.024211298020746064
|
| 947 |
+
},
|
| 948 |
+
{
|
| 949 |
+
"run_mode": "locked_stream",
|
| 950 |
+
"condition": "static_dropout_0.02",
|
| 951 |
+
"condition_kind": "static",
|
| 952 |
+
"stage": 3,
|
| 953 |
+
"token_limit": 2000000,
|
| 954 |
+
"model_name": "L12_H8_D320",
|
| 955 |
+
"n_layer": 12,
|
| 956 |
+
"n_head": 8,
|
| 957 |
+
"n_embd": 320,
|
| 958 |
+
"parameters": 17367040,
|
| 959 |
+
"dropout_initial": 0.02,
|
| 960 |
+
"dropout_final": 0.02,
|
| 961 |
+
"dropout_schedule": "constant",
|
| 962 |
+
"n": 5,
|
| 963 |
+
"mean_train_eval_loss": 3.6672953754663467,
|
| 964 |
+
"std_train_eval_loss": 0.010966930251295475,
|
| 965 |
+
"mean_val_eval_loss": 4.355083760619164,
|
| 966 |
+
"std_val_eval_loss": 0.026489436313039863,
|
| 967 |
+
"mean_generalization_gap": 0.6877883851528168,
|
| 968 |
+
"std_generalization_gap": 0.026106538952549097
|
| 969 |
+
},
|
| 970 |
+
{
|
| 971 |
+
"run_mode": "locked_stream",
|
| 972 |
+
"condition": "static_dropout_0.04",
|
| 973 |
+
"condition_kind": "static",
|
| 974 |
+
"stage": 3,
|
| 975 |
+
"token_limit": 2000000,
|
| 976 |
+
"model_name": "L12_H8_D320",
|
| 977 |
+
"n_layer": 12,
|
| 978 |
+
"n_head": 8,
|
| 979 |
+
"n_embd": 320,
|
| 980 |
+
"parameters": 17367040,
|
| 981 |
+
"dropout_initial": 0.04,
|
| 982 |
+
"dropout_final": 0.04,
|
| 983 |
+
"dropout_schedule": "constant",
|
| 984 |
+
"n": 5,
|
| 985 |
+
"mean_train_eval_loss": 3.703781445324421,
|
| 986 |
+
"std_train_eval_loss": 0.01970997472996272,
|
| 987 |
+
"mean_val_eval_loss": 4.3262467920780185,
|
| 988 |
+
"std_val_eval_loss": 0.016533384756725625,
|
| 989 |
+
"mean_generalization_gap": 0.6224653467535972,
|
| 990 |
+
"std_generalization_gap": 0.02577206799431021
|
| 991 |
+
},
|
| 992 |
+
{
|
| 993 |
+
"run_mode": "locked_stream",
|
| 994 |
+
"condition": "static_dropout_0.06",
|
| 995 |
+
"condition_kind": "static",
|
| 996 |
+
"stage": 3,
|
| 997 |
+
"token_limit": 2000000,
|
| 998 |
+
"model_name": "L12_H8_D320",
|
| 999 |
+
"n_layer": 12,
|
| 1000 |
+
"n_head": 8,
|
| 1001 |
+
"n_embd": 320,
|
| 1002 |
+
"parameters": 17367040,
|
| 1003 |
+
"dropout_initial": 0.06,
|
| 1004 |
+
"dropout_final": 0.06,
|
| 1005 |
+
"dropout_schedule": "constant",
|
| 1006 |
+
"n": 5,
|
| 1007 |
+
"mean_train_eval_loss": 3.7317687764763834,
|
| 1008 |
+
"std_train_eval_loss": 0.017157479537739013,
|
| 1009 |
+
"mean_val_eval_loss": 4.305900506675243,
|
| 1010 |
+
"std_val_eval_loss": 0.018234543669842573,
|
| 1011 |
+
"mean_generalization_gap": 0.5741317301988602,
|
| 1012 |
+
"std_generalization_gap": 0.023633182905307998
|
| 1013 |
+
},
|
| 1014 |
+
{
|
| 1015 |
+
"run_mode": "locked_stream",
|
| 1016 |
+
"condition": "static_dropout_0.08",
|
| 1017 |
+
"condition_kind": "static",
|
| 1018 |
+
"stage": 3,
|
| 1019 |
+
"token_limit": 2000000,
|
| 1020 |
+
"model_name": "L12_H8_D320",
|
| 1021 |
+
"n_layer": 12,
|
| 1022 |
+
"n_head": 8,
|
| 1023 |
+
"n_embd": 320,
|
| 1024 |
+
"parameters": 17367040,
|
| 1025 |
+
"dropout_initial": 0.08,
|
| 1026 |
+
"dropout_final": 0.08,
|
| 1027 |
+
"dropout_schedule": "constant",
|
| 1028 |
+
"n": 5,
|
| 1029 |
+
"mean_train_eval_loss": 3.7518652200698854,
|
| 1030 |
+
"std_train_eval_loss": 0.019831550751639896,
|
| 1031 |
+
"mean_val_eval_loss": 4.292461846768856,
|
| 1032 |
+
"std_val_eval_loss": 0.016558424281915483,
|
| 1033 |
+
"mean_generalization_gap": 0.5405966266989708,
|
| 1034 |
+
"std_generalization_gap": 0.034040566494281235
|
| 1035 |
+
},
|
| 1036 |
+
{
|
| 1037 |
+
"run_mode": "locked_stream",
|
| 1038 |
+
"condition": "static_dropout_0.1",
|
| 1039 |
+
"condition_kind": "static",
|
| 1040 |
+
"stage": 3,
|
| 1041 |
+
"token_limit": 2000000,
|
| 1042 |
+
"model_name": "L12_H8_D320",
|
| 1043 |
+
"n_layer": 12,
|
| 1044 |
+
"n_head": 8,
|
| 1045 |
+
"n_embd": 320,
|
| 1046 |
+
"parameters": 17367040,
|
| 1047 |
+
"dropout_initial": 0.1,
|
| 1048 |
+
"dropout_final": 0.1,
|
| 1049 |
+
"dropout_schedule": "constant",
|
| 1050 |
+
"n": 5,
|
| 1051 |
+
"mean_train_eval_loss": 3.7808884128928186,
|
| 1052 |
+
"std_train_eval_loss": 0.01766528157534878,
|
| 1053 |
+
"mean_val_eval_loss": 4.285703422129155,
|
| 1054 |
+
"std_val_eval_loss": 0.018994111876547076,
|
| 1055 |
+
"mean_generalization_gap": 0.5048150092363357,
|
| 1056 |
+
"std_generalization_gap": 0.028072111875115907
|
| 1057 |
+
},
|
| 1058 |
+
{
|
| 1059 |
+
"run_mode": "locked_stream",
|
| 1060 |
+
"condition": "static_dropout_0.14",
|
| 1061 |
+
"condition_kind": "static",
|
| 1062 |
+
"stage": 3,
|
| 1063 |
+
"token_limit": 2000000,
|
| 1064 |
+
"model_name": "L12_H8_D320",
|
| 1065 |
+
"n_layer": 12,
|
| 1066 |
+
"n_head": 8,
|
| 1067 |
+
"n_embd": 320,
|
| 1068 |
+
"parameters": 17367040,
|
| 1069 |
+
"dropout_initial": 0.14,
|
| 1070 |
+
"dropout_final": 0.14,
|
| 1071 |
+
"dropout_schedule": "constant",
|
| 1072 |
+
"n": 5,
|
| 1073 |
+
"mean_train_eval_loss": 3.828673002123833,
|
| 1074 |
+
"std_train_eval_loss": 0.004937581309037719,
|
| 1075 |
+
"mean_val_eval_loss": 4.281563547253609,
|
| 1076 |
+
"std_val_eval_loss": 0.02138858846966249,
|
| 1077 |
+
"mean_generalization_gap": 0.452890545129776,
|
| 1078 |
+
"std_generalization_gap": 0.01755457638984346
|
| 1079 |
+
},
|
| 1080 |
+
{
|
| 1081 |
+
"run_mode": "locked_stream",
|
| 1082 |
+
"condition": "static_dropout_0.18",
|
| 1083 |
+
"condition_kind": "static",
|
| 1084 |
+
"stage": 3,
|
| 1085 |
+
"token_limit": 2000000,
|
| 1086 |
+
"model_name": "L12_H8_D320",
|
| 1087 |
+
"n_layer": 12,
|
| 1088 |
+
"n_head": 8,
|
| 1089 |
+
"n_embd": 320,
|
| 1090 |
+
"parameters": 17367040,
|
| 1091 |
+
"dropout_initial": 0.18,
|
| 1092 |
+
"dropout_final": 0.18,
|
| 1093 |
+
"dropout_schedule": "constant",
|
| 1094 |
+
"n": 5,
|
| 1095 |
+
"mean_train_eval_loss": 3.876796562969685,
|
| 1096 |
+
"std_train_eval_loss": 0.008357602777023191,
|
| 1097 |
+
"mean_val_eval_loss": 4.288915157318115,
|
| 1098 |
+
"std_val_eval_loss": 0.019532110178580833,
|
| 1099 |
+
"mean_generalization_gap": 0.4121185943484306,
|
| 1100 |
+
"std_generalization_gap": 0.013904819825939841
|
| 1101 |
+
},
|
| 1102 |
+
{
|
| 1103 |
+
"run_mode": "locked_stream",
|
| 1104 |
+
"condition": "static_dropout_0.2",
|
| 1105 |
+
"condition_kind": "static",
|
| 1106 |
+
"stage": 3,
|
| 1107 |
+
"token_limit": 2000000,
|
| 1108 |
+
"model_name": "L12_H8_D320",
|
| 1109 |
+
"n_layer": 12,
|
| 1110 |
+
"n_head": 8,
|
| 1111 |
+
"n_embd": 320,
|
| 1112 |
+
"parameters": 17367040,
|
| 1113 |
+
"dropout_initial": 0.2,
|
| 1114 |
+
"dropout_final": 0.2,
|
| 1115 |
+
"dropout_schedule": "constant",
|
| 1116 |
+
"n": 5,
|
| 1117 |
+
"mean_train_eval_loss": 3.898232290148735,
|
| 1118 |
+
"std_train_eval_loss": 0.008893270752025355,
|
| 1119 |
+
"mean_val_eval_loss": 4.294206875562668,
|
| 1120 |
+
"std_val_eval_loss": 0.016019909189912813,
|
| 1121 |
+
"mean_generalization_gap": 0.3959745854139328,
|
| 1122 |
+
"std_generalization_gap": 0.01268490098888804
|
| 1123 |
+
},
|
| 1124 |
+
{
|
| 1125 |
+
"run_mode": "locked_stream",
|
| 1126 |
+
"condition": "static_dropout_0.26",
|
| 1127 |
+
"condition_kind": "static",
|
| 1128 |
+
"stage": 3,
|
| 1129 |
+
"token_limit": 2000000,
|
| 1130 |
+
"model_name": "L12_H8_D320",
|
| 1131 |
+
"n_layer": 12,
|
| 1132 |
+
"n_head": 8,
|
| 1133 |
+
"n_embd": 320,
|
| 1134 |
+
"parameters": 17367040,
|
| 1135 |
+
"dropout_initial": 0.26,
|
| 1136 |
+
"dropout_final": 0.26,
|
| 1137 |
+
"dropout_schedule": "constant",
|
| 1138 |
+
"n": 5,
|
| 1139 |
+
"mean_train_eval_loss": 3.9654731526970863,
|
| 1140 |
+
"std_train_eval_loss": 0.018160558354021094,
|
| 1141 |
+
"mean_val_eval_loss": 4.32477528527379,
|
| 1142 |
+
"std_val_eval_loss": 0.017871972287011016,
|
| 1143 |
+
"mean_generalization_gap": 0.359302132576704,
|
| 1144 |
+
"std_generalization_gap": 0.018728321745075875
|
| 1145 |
+
},
|
| 1146 |
+
{
|
| 1147 |
+
"run_mode": "locked_stream",
|
| 1148 |
+
"condition": "static_dropout_0.3",
|
| 1149 |
+
"condition_kind": "static",
|
| 1150 |
+
"stage": 3,
|
| 1151 |
+
"token_limit": 2000000,
|
| 1152 |
+
"model_name": "L12_H8_D320",
|
| 1153 |
+
"n_layer": 12,
|
| 1154 |
+
"n_head": 8,
|
| 1155 |
+
"n_embd": 320,
|
| 1156 |
+
"parameters": 17367040,
|
| 1157 |
+
"dropout_initial": 0.3,
|
| 1158 |
+
"dropout_final": 0.3,
|
| 1159 |
+
"dropout_schedule": "constant",
|
| 1160 |
+
"n": 5,
|
| 1161 |
+
"mean_train_eval_loss": 4.009651578962803,
|
| 1162 |
+
"std_train_eval_loss": 0.014639843511797184,
|
| 1163 |
+
"mean_val_eval_loss": 4.34667577445507,
|
| 1164 |
+
"std_val_eval_loss": 0.015854348211857338,
|
| 1165 |
+
"mean_generalization_gap": 0.3370241954922676,
|
| 1166 |
+
"std_generalization_gap": 0.019075525784314876
|
| 1167 |
+
},
|
| 1168 |
+
{
|
| 1169 |
+
"run_mode": "locked_stream",
|
| 1170 |
+
"condition": "wikitext103_formula_l12",
|
| 1171 |
+
"condition_kind": "anchor_decay",
|
| 1172 |
+
"stage": 3,
|
| 1173 |
+
"token_limit": 2000000,
|
| 1174 |
+
"model_name": "L12_H8_D320",
|
| 1175 |
+
"n_layer": 12,
|
| 1176 |
+
"n_head": 8,
|
| 1177 |
+
"n_embd": 320,
|
| 1178 |
+
"parameters": 17367040,
|
| 1179 |
+
"dropout_initial": 0.3,
|
| 1180 |
+
"dropout_final": 0.02,
|
| 1181 |
+
"dropout_schedule": "log_prefix_anchor",
|
| 1182 |
+
"n": 5,
|
| 1183 |
+
"mean_train_eval_loss": 3.8088956013321877,
|
| 1184 |
+
"std_train_eval_loss": 0.016376701234484724,
|
| 1185 |
+
"mean_val_eval_loss": 4.260725708305836,
|
| 1186 |
+
"std_val_eval_loss": 0.018098739272622868,
|
| 1187 |
+
"mean_generalization_gap": 0.45183010697364806,
|
| 1188 |
+
"std_generalization_gap": 0.01960955973506382
|
| 1189 |
+
},
|
| 1190 |
+
{
|
| 1191 |
+
"run_mode": "locked_stream",
|
| 1192 |
+
"condition": "wikitext103_low_decay",
|
| 1193 |
+
"condition_kind": "anchor_decay",
|
| 1194 |
+
"stage": 3,
|
| 1195 |
+
"token_limit": 2000000,
|
| 1196 |
+
"model_name": "L12_H8_D320",
|
| 1197 |
+
"n_layer": 12,
|
| 1198 |
+
"n_head": 8,
|
| 1199 |
+
"n_embd": 320,
|
| 1200 |
+
"parameters": 17367040,
|
| 1201 |
+
"dropout_initial": 0.14,
|
| 1202 |
+
"dropout_final": 0.02,
|
| 1203 |
+
"dropout_schedule": "log_prefix_anchor",
|
| 1204 |
+
"n": 5,
|
| 1205 |
+
"mean_train_eval_loss": 3.7474397838115694,
|
| 1206 |
+
"std_train_eval_loss": 0.003830485352807811,
|
| 1207 |
+
"mean_val_eval_loss": 4.273568108677864,
|
| 1208 |
+
"std_val_eval_loss": 0.022770462958788148,
|
| 1209 |
+
"mean_generalization_gap": 0.5261283248662949,
|
| 1210 |
+
"std_generalization_gap": 0.02051193157067845
|
| 1211 |
+
},
|
| 1212 |
+
{
|
| 1213 |
+
"run_mode": "locked_stream",
|
| 1214 |
+
"condition": "wikitext103_probe_blend",
|
| 1215 |
+
"condition_kind": "anchor_decay",
|
| 1216 |
+
"stage": 3,
|
| 1217 |
+
"token_limit": 2000000,
|
| 1218 |
+
"model_name": "L12_H8_D320",
|
| 1219 |
+
"n_layer": 12,
|
| 1220 |
+
"n_head": 8,
|
| 1221 |
+
"n_embd": 320,
|
| 1222 |
+
"parameters": 17367040,
|
| 1223 |
+
"dropout_initial": 0.19,
|
| 1224 |
+
"dropout_final": 0.01,
|
| 1225 |
+
"dropout_schedule": "log_prefix_anchor",
|
| 1226 |
+
"n": 5,
|
| 1227 |
+
"mean_train_eval_loss": 3.720025636255741,
|
| 1228 |
+
"std_train_eval_loss": 0.012889615922098394,
|
| 1229 |
+
"mean_val_eval_loss": 4.274319227784872,
|
| 1230 |
+
"std_val_eval_loss": 0.0173783813582558,
|
| 1231 |
+
"mean_generalization_gap": 0.5542935915291309,
|
| 1232 |
+
"std_generalization_gap": 0.012746502538982045
|
| 1233 |
+
},
|
| 1234 |
+
{
|
| 1235 |
+
"run_mode": "locked_stream",
|
| 1236 |
+
"condition": "static_dropout_0",
|
| 1237 |
+
"condition_kind": "static",
|
| 1238 |
+
"stage": 4,
|
| 1239 |
+
"token_limit": 4000000,
|
| 1240 |
+
"model_name": "L12_H8_D320",
|
| 1241 |
+
"n_layer": 12,
|
| 1242 |
+
"n_head": 8,
|
| 1243 |
+
"n_embd": 320,
|
| 1244 |
+
"parameters": 17367040,
|
| 1245 |
+
"dropout_initial": 0.0,
|
| 1246 |
+
"dropout_final": 0.0,
|
| 1247 |
+
"dropout_schedule": "constant",
|
| 1248 |
+
"n": 5,
|
| 1249 |
+
"mean_train_eval_loss": 3.7750062167644503,
|
| 1250 |
+
"std_train_eval_loss": 0.025453964920340872,
|
| 1251 |
+
"mean_val_eval_loss": 4.183486974239349,
|
| 1252 |
+
"std_val_eval_loss": 0.01652781602088682,
|
| 1253 |
+
"mean_generalization_gap": 0.4084807574748993,
|
| 1254 |
+
"std_generalization_gap": 0.03358628463770181
|
| 1255 |
+
},
|
| 1256 |
+
{
|
| 1257 |
+
"run_mode": "locked_stream",
|
| 1258 |
+
"condition": "static_dropout_0.02",
|
| 1259 |
+
"condition_kind": "static",
|
| 1260 |
+
"stage": 4,
|
| 1261 |
+
"token_limit": 4000000,
|
| 1262 |
+
"model_name": "L12_H8_D320",
|
| 1263 |
+
"n_layer": 12,
|
| 1264 |
+
"n_head": 8,
|
| 1265 |
+
"n_embd": 320,
|
| 1266 |
+
"parameters": 17367040,
|
| 1267 |
+
"dropout_initial": 0.02,
|
| 1268 |
+
"dropout_final": 0.02,
|
| 1269 |
+
"dropout_schedule": "constant",
|
| 1270 |
+
"n": 5,
|
| 1271 |
+
"mean_train_eval_loss": 3.775918734073639,
|
| 1272 |
+
"std_train_eval_loss": 0.03360785862475549,
|
| 1273 |
+
"mean_val_eval_loss": 4.145870254188776,
|
| 1274 |
+
"std_val_eval_loss": 0.016455755256661365,
|
| 1275 |
+
"mean_generalization_gap": 0.3699515201151371,
|
| 1276 |
+
"std_generalization_gap": 0.0368559849747316
|
| 1277 |
+
},
|
| 1278 |
+
{
|
| 1279 |
+
"run_mode": "locked_stream",
|
| 1280 |
+
"condition": "static_dropout_0.04",
|
| 1281 |
+
"condition_kind": "static",
|
| 1282 |
+
"stage": 4,
|
| 1283 |
+
"token_limit": 4000000,
|
| 1284 |
+
"model_name": "L12_H8_D320",
|
| 1285 |
+
"n_layer": 12,
|
| 1286 |
+
"n_head": 8,
|
| 1287 |
+
"n_embd": 320,
|
| 1288 |
+
"parameters": 17367040,
|
| 1289 |
+
"dropout_initial": 0.04,
|
| 1290 |
+
"dropout_final": 0.04,
|
| 1291 |
+
"dropout_schedule": "constant",
|
| 1292 |
+
"n": 5,
|
| 1293 |
+
"mean_train_eval_loss": 3.7978432700037956,
|
| 1294 |
+
"std_train_eval_loss": 0.018504253509089505,
|
| 1295 |
+
"mean_val_eval_loss": 4.1331265024840835,
|
| 1296 |
+
"std_val_eval_loss": 0.022679096058077455,
|
| 1297 |
+
"mean_generalization_gap": 0.3352832324802876,
|
| 1298 |
+
"std_generalization_gap": 0.036956450676459744
|
| 1299 |
+
},
|
| 1300 |
+
{
|
| 1301 |
+
"run_mode": "locked_stream",
|
| 1302 |
+
"condition": "static_dropout_0.06",
|
| 1303 |
+
"condition_kind": "static",
|
| 1304 |
+
"stage": 4,
|
| 1305 |
+
"token_limit": 4000000,
|
| 1306 |
+
"model_name": "L12_H8_D320",
|
| 1307 |
+
"n_layer": 12,
|
| 1308 |
+
"n_head": 8,
|
| 1309 |
+
"n_embd": 320,
|
| 1310 |
+
"parameters": 17367040,
|
| 1311 |
+
"dropout_initial": 0.06,
|
| 1312 |
+
"dropout_final": 0.06,
|
| 1313 |
+
"dropout_schedule": "constant",
|
| 1314 |
+
"n": 5,
|
| 1315 |
+
"mean_train_eval_loss": 3.806561966240406,
|
| 1316 |
+
"std_train_eval_loss": 0.022924307415498075,
|
| 1317 |
+
"mean_val_eval_loss": 4.119682380557061,
|
| 1318 |
+
"std_val_eval_loss": 0.008179605355252681,
|
| 1319 |
+
"mean_generalization_gap": 0.3131204143166542,
|
| 1320 |
+
"std_generalization_gap": 0.028667268842643943
|
| 1321 |
+
},
|
| 1322 |
+
{
|
| 1323 |
+
"run_mode": "locked_stream",
|
| 1324 |
+
"condition": "static_dropout_0.08",
|
| 1325 |
+
"condition_kind": "static",
|
| 1326 |
+
"stage": 4,
|
| 1327 |
+
"token_limit": 4000000,
|
| 1328 |
+
"model_name": "L12_H8_D320",
|
| 1329 |
+
"n_layer": 12,
|
| 1330 |
+
"n_head": 8,
|
| 1331 |
+
"n_embd": 320,
|
| 1332 |
+
"parameters": 17367040,
|
| 1333 |
+
"dropout_initial": 0.08,
|
| 1334 |
+
"dropout_final": 0.08,
|
| 1335 |
+
"dropout_schedule": "constant",
|
| 1336 |
+
"n": 5,
|
| 1337 |
+
"mean_train_eval_loss": 3.8268220275640488,
|
| 1338 |
+
"std_train_eval_loss": 0.021203534366681782,
|
| 1339 |
+
"mean_val_eval_loss": 4.111632210761309,
|
| 1340 |
+
"std_val_eval_loss": 0.01857117977689519,
|
| 1341 |
+
"mean_generalization_gap": 0.2848101831972599,
|
| 1342 |
+
"std_generalization_gap": 0.03679859980086126
|
| 1343 |
+
},
|
| 1344 |
+
{
|
| 1345 |
+
"run_mode": "locked_stream",
|
| 1346 |
+
"condition": "static_dropout_0.1",
|
| 1347 |
+
"condition_kind": "static",
|
| 1348 |
+
"stage": 4,
|
| 1349 |
+
"token_limit": 4000000,
|
| 1350 |
+
"model_name": "L12_H8_D320",
|
| 1351 |
+
"n_layer": 12,
|
| 1352 |
+
"n_head": 8,
|
| 1353 |
+
"n_embd": 320,
|
| 1354 |
+
"parameters": 17367040,
|
| 1355 |
+
"dropout_initial": 0.1,
|
| 1356 |
+
"dropout_final": 0.1,
|
| 1357 |
+
"dropout_schedule": "constant",
|
| 1358 |
+
"n": 5,
|
| 1359 |
+
"mean_train_eval_loss": 3.841722105443478,
|
| 1360 |
+
"std_train_eval_loss": 0.019884552697005082,
|
| 1361 |
+
"mean_val_eval_loss": 4.1104619488120075,
|
| 1362 |
+
"std_val_eval_loss": 0.018806640195540816,
|
| 1363 |
+
"mean_generalization_gap": 0.2687398433685303,
|
| 1364 |
+
"std_generalization_gap": 0.03219305563944893
|
| 1365 |
+
},
|
| 1366 |
+
{
|
| 1367 |
+
"run_mode": "locked_stream",
|
| 1368 |
+
"condition": "static_dropout_0.14",
|
| 1369 |
+
"condition_kind": "static",
|
| 1370 |
+
"stage": 4,
|
| 1371 |
+
"token_limit": 4000000,
|
| 1372 |
+
"model_name": "L12_H8_D320",
|
| 1373 |
+
"n_layer": 12,
|
| 1374 |
+
"n_head": 8,
|
| 1375 |
+
"n_embd": 320,
|
| 1376 |
+
"parameters": 17367040,
|
| 1377 |
+
"dropout_initial": 0.14,
|
| 1378 |
+
"dropout_final": 0.14,
|
| 1379 |
+
"dropout_schedule": "constant",
|
| 1380 |
+
"n": 5,
|
| 1381 |
+
"mean_train_eval_loss": 3.8673695802688597,
|
| 1382 |
+
"std_train_eval_loss": 0.025363134860510252,
|
| 1383 |
+
"mean_val_eval_loss": 4.122133788466454,
|
| 1384 |
+
"std_val_eval_loss": 0.015529540565590345,
|
| 1385 |
+
"mean_generalization_gap": 0.2547642081975937,
|
| 1386 |
+
"std_generalization_gap": 0.0302072467927019
|
| 1387 |
+
},
|
| 1388 |
+
{
|
| 1389 |
+
"run_mode": "locked_stream",
|
| 1390 |
+
"condition": "static_dropout_0.18",
|
| 1391 |
+
"condition_kind": "static",
|
| 1392 |
+
"stage": 4,
|
| 1393 |
+
"token_limit": 4000000,
|
| 1394 |
+
"model_name": "L12_H8_D320",
|
| 1395 |
+
"n_layer": 12,
|
| 1396 |
+
"n_head": 8,
|
| 1397 |
+
"n_embd": 320,
|
| 1398 |
+
"parameters": 17367040,
|
| 1399 |
+
"dropout_initial": 0.18,
|
| 1400 |
+
"dropout_final": 0.18,
|
| 1401 |
+
"dropout_schedule": "constant",
|
| 1402 |
+
"n": 5,
|
| 1403 |
+
"mean_train_eval_loss": 3.9014849200844766,
|
| 1404 |
+
"std_train_eval_loss": 0.018144303418082167,
|
| 1405 |
+
"mean_val_eval_loss": 4.130411610752344,
|
| 1406 |
+
"std_val_eval_loss": 0.013005735782293457,
|
| 1407 |
+
"mean_generalization_gap": 0.22892669066786767,
|
| 1408 |
+
"std_generalization_gap": 0.02966442514236336
|
| 1409 |
+
},
|
| 1410 |
+
{
|
| 1411 |
+
"run_mode": "locked_stream",
|
| 1412 |
+
"condition": "static_dropout_0.2",
|
| 1413 |
+
"condition_kind": "static",
|
| 1414 |
+
"stage": 4,
|
| 1415 |
+
"token_limit": 4000000,
|
| 1416 |
+
"model_name": "L12_H8_D320",
|
| 1417 |
+
"n_layer": 12,
|
| 1418 |
+
"n_head": 8,
|
| 1419 |
+
"n_embd": 320,
|
| 1420 |
+
"parameters": 17367040,
|
| 1421 |
+
"dropout_initial": 0.2,
|
| 1422 |
+
"dropout_final": 0.2,
|
| 1423 |
+
"dropout_schedule": "constant",
|
| 1424 |
+
"n": 5,
|
| 1425 |
+
"mean_train_eval_loss": 3.9154964342713354,
|
| 1426 |
+
"std_train_eval_loss": 0.020116706552102667,
|
| 1427 |
+
"mean_val_eval_loss": 4.1393771633505825,
|
| 1428 |
+
"std_val_eval_loss": 0.016742174044810716,
|
| 1429 |
+
"mean_generalization_gap": 0.22388072907924653,
|
| 1430 |
+
"std_generalization_gap": 0.033233695598367155
|
| 1431 |
+
},
|
| 1432 |
+
{
|
| 1433 |
+
"run_mode": "locked_stream",
|
| 1434 |
+
"condition": "static_dropout_0.26",
|
| 1435 |
+
"condition_kind": "static",
|
| 1436 |
+
"stage": 4,
|
| 1437 |
+
"token_limit": 4000000,
|
| 1438 |
+
"model_name": "L12_H8_D320",
|
| 1439 |
+
"n_layer": 12,
|
| 1440 |
+
"n_head": 8,
|
| 1441 |
+
"n_embd": 320,
|
| 1442 |
+
"parameters": 17367040,
|
| 1443 |
+
"dropout_initial": 0.26,
|
| 1444 |
+
"dropout_final": 0.26,
|
| 1445 |
+
"dropout_schedule": "constant",
|
| 1446 |
+
"n": 5,
|
| 1447 |
+
"mean_train_eval_loss": 3.9775424867868425,
|
| 1448 |
+
"std_train_eval_loss": 0.021977989750098077,
|
| 1449 |
+
"mean_val_eval_loss": 4.178351600468159,
|
| 1450 |
+
"std_val_eval_loss": 0.01453281035529064,
|
| 1451 |
+
"mean_generalization_gap": 0.20080911368131638,
|
| 1452 |
+
"std_generalization_gap": 0.03295020359909027
|
| 1453 |
+
},
|
| 1454 |
+
{
|
| 1455 |
+
"run_mode": "locked_stream",
|
| 1456 |
+
"condition": "static_dropout_0.3",
|
| 1457 |
+
"condition_kind": "static",
|
| 1458 |
+
"stage": 4,
|
| 1459 |
+
"token_limit": 4000000,
|
| 1460 |
+
"model_name": "L12_H8_D320",
|
| 1461 |
+
"n_layer": 12,
|
| 1462 |
+
"n_head": 8,
|
| 1463 |
+
"n_embd": 320,
|
| 1464 |
+
"parameters": 17367040,
|
| 1465 |
+
"dropout_initial": 0.3,
|
| 1466 |
+
"dropout_final": 0.3,
|
| 1467 |
+
"dropout_schedule": "constant",
|
| 1468 |
+
"n": 5,
|
| 1469 |
+
"mean_train_eval_loss": 4.012722708284855,
|
| 1470 |
+
"std_train_eval_loss": 0.024177126067976527,
|
| 1471 |
+
"mean_val_eval_loss": 4.194609892368317,
|
| 1472 |
+
"std_val_eval_loss": 0.014131822174274241,
|
| 1473 |
+
"mean_generalization_gap": 0.18188718408346177,
|
| 1474 |
+
"std_generalization_gap": 0.031100561007876955
|
| 1475 |
+
},
|
| 1476 |
+
{
|
| 1477 |
+
"run_mode": "locked_stream",
|
| 1478 |
+
"condition": "wikitext103_formula_l12",
|
| 1479 |
+
"condition_kind": "anchor_decay",
|
| 1480 |
+
"stage": 4,
|
| 1481 |
+
"token_limit": 4000000,
|
| 1482 |
+
"model_name": "L12_H8_D320",
|
| 1483 |
+
"n_layer": 12,
|
| 1484 |
+
"n_head": 8,
|
| 1485 |
+
"n_embd": 320,
|
| 1486 |
+
"parameters": 17367040,
|
| 1487 |
+
"dropout_initial": 0.3,
|
| 1488 |
+
"dropout_final": 0.02,
|
| 1489 |
+
"dropout_schedule": "log_prefix_anchor",
|
| 1490 |
+
"n": 5,
|
| 1491 |
+
"mean_train_eval_loss": 3.7991417959332465,
|
| 1492 |
+
"std_train_eval_loss": 0.028273745467366926,
|
| 1493 |
+
"mean_val_eval_loss": 4.080797865241766,
|
| 1494 |
+
"std_val_eval_loss": 0.019458062276034267,
|
| 1495 |
+
"mean_generalization_gap": 0.2816560693085194,
|
| 1496 |
+
"std_generalization_gap": 0.038849358858741
|
| 1497 |
+
},
|
| 1498 |
+
{
|
| 1499 |
+
"run_mode": "locked_stream",
|
| 1500 |
+
"condition": "wikitext103_low_decay",
|
| 1501 |
+
"condition_kind": "anchor_decay",
|
| 1502 |
+
"stage": 4,
|
| 1503 |
+
"token_limit": 4000000,
|
| 1504 |
+
"model_name": "L12_H8_D320",
|
| 1505 |
+
"n_layer": 12,
|
| 1506 |
+
"n_head": 8,
|
| 1507 |
+
"n_embd": 320,
|
| 1508 |
+
"parameters": 17367040,
|
| 1509 |
+
"dropout_initial": 0.14,
|
| 1510 |
+
"dropout_final": 0.02,
|
| 1511 |
+
"dropout_schedule": "log_prefix_anchor",
|
| 1512 |
+
"n": 5,
|
| 1513 |
+
"mean_train_eval_loss": 3.776866267621517,
|
| 1514 |
+
"std_train_eval_loss": 0.02647416884246062,
|
| 1515 |
+
"mean_val_eval_loss": 4.102013133466244,
|
| 1516 |
+
"std_val_eval_loss": 0.016554390946653275,
|
| 1517 |
+
"mean_generalization_gap": 0.3251468658447266,
|
| 1518 |
+
"std_generalization_gap": 0.03314100249583806
|
| 1519 |
+
},
|
| 1520 |
+
{
|
| 1521 |
+
"run_mode": "locked_stream",
|
| 1522 |
+
"condition": "wikitext103_probe_blend",
|
| 1523 |
+
"condition_kind": "anchor_decay",
|
| 1524 |
+
"stage": 4,
|
| 1525 |
+
"token_limit": 4000000,
|
| 1526 |
+
"model_name": "L12_H8_D320",
|
| 1527 |
+
"n_layer": 12,
|
| 1528 |
+
"n_head": 8,
|
| 1529 |
+
"n_embd": 320,
|
| 1530 |
+
"parameters": 17367040,
|
| 1531 |
+
"dropout_initial": 0.19,
|
| 1532 |
+
"dropout_final": 0.01,
|
| 1533 |
+
"dropout_schedule": "log_prefix_anchor",
|
| 1534 |
+
"n": 5,
|
| 1535 |
+
"mean_train_eval_loss": 3.7674024373292925,
|
| 1536 |
+
"std_train_eval_loss": 0.027214408565503653,
|
| 1537 |
+
"mean_val_eval_loss": 4.096089626103639,
|
| 1538 |
+
"std_val_eval_loss": 0.014523730760550201,
|
| 1539 |
+
"mean_generalization_gap": 0.3286871887743473,
|
| 1540 |
+
"std_generalization_gap": 0.038218075738052845
|
| 1541 |
+
}
|
| 1542 |
+
]
|
runs/wikitext103_l12_streaming_validation_5seed/locked_stream/20260531-093525/trace.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
runs/wikitext103_streaming_report/l12_validation_5seed/condition_summary.csv
ADDED
|
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
condition,kind,n,mean_trajectory_val,std_trajectory_val,mean_final_val,std_final_val,mean_final_gap,std_final_gap,dropout_path
|
| 2 |
+
wikitext103_formula_l12,anchor_decay,5,4.571096663624049,0.004493905823534624,4.080797865241766,0.019458062276034267,0.2816560693085194,0.038849358858741,0.30 -> 0.26 -> 0.18 -> 0.09 -> 0.02
|
| 3 |
+
wikitext103_probe_blend,anchor_decay,5,4.5635083836317065,0.0046223860048219215,4.096089626103639,0.014523730760550201,0.3286871887743473,0.038218075738052845,0.19 -> 0.14 -> 0.09 -> 0.04 -> 0.01
|
| 4 |
+
wikitext103_low_decay,anchor_decay,5,4.568109618723392,0.007251133716114802,4.102013133466244,0.016554390946653275,0.3251468658447266,0.03314100249583806,0.14 -> 0.14 -> 0.10 -> 0.06 -> 0.02
|
| 5 |
+
static_dropout_0.1,static,5,4.583564288020134,0.00624825224116908,4.1104619488120075,0.018806640195540816,0.2687398433685303,0.03219305563944893,0.10 -> 0.10 -> 0.10 -> 0.10 -> 0.10
|
| 6 |
+
static_dropout_0.08,static,5,4.596727982312441,0.007294540304515209,4.111632210761309,0.01857117977689519,0.2848101831972599,0.03679859980086126,0.08 -> 0.08 -> 0.08 -> 0.08 -> 0.08
|
| 7 |
+
static_dropout_0.06,static,5,4.618606640398502,0.004793067320536325,4.119682380557061,0.008179605355252681,0.3131204143166542,0.028667268842643943,0.06 -> 0.06 -> 0.06 -> 0.06 -> 0.06
|
| 8 |
+
static_dropout_0.14,static,5,4.573477408885956,0.007705324015227247,4.122133788466454,0.015529540565590345,0.2547642081975937,0.0302072467927019,0.14 -> 0.14 -> 0.14 -> 0.14 -> 0.14
|
| 9 |
+
static_dropout_0.18,static,5,4.575578650385141,0.004087884073210763,4.130411610752344,0.013005735782293457,0.22892669066786767,0.02966442514236336,0.18 -> 0.18 -> 0.18 -> 0.18 -> 0.18
|
| 10 |
+
static_dropout_0.04,static,5,4.6500934390723705,0.007686314739184788,4.1331265024840835,0.022679096058077455,0.3352832324802876,0.036956450676459744,0.04 -> 0.04 -> 0.04 -> 0.04 -> 0.04
|
| 11 |
+
static_dropout_0.2,static,5,4.579393386542797,0.004961083585861927,4.1393771633505825,0.016742174044810716,0.22388072907924653,0.033233695598367155,0.20 -> 0.20 -> 0.20 -> 0.20 -> 0.20
|
| 12 |
+
static_dropout_0.02,static,5,4.695371071547269,0.008590324433750853,4.145870254188776,0.016455755256661365,0.3699515201151371,0.0368559849747316,0.02 -> 0.02 -> 0.02 -> 0.02 -> 0.02
|
| 13 |
+
static_dropout_0.26,static,5,4.606293898075819,0.0051111340096229975,4.178351600468159,0.01453281035529064,0.20080911368131638,0.03295020359909027,0.26 -> 0.26 -> 0.26 -> 0.26 -> 0.26
|
| 14 |
+
static_dropout_0,static,5,4.776249321401119,0.010887492470564343,4.183486974239349,0.01652781602088682,0.4084807574748993,0.03358628463770181,0.00 -> 0.00 -> 0.00 -> 0.00 -> 0.00
|
| 15 |
+
static_dropout_0.3,static,5,4.625303542017937,0.003399857155492287,4.194609892368317,0.014131822174274241,0.18188718408346177,0.031100561007876955,0.30 -> 0.30 -> 0.30 -> 0.30 -> 0.30
|
runs/wikitext103_streaming_report/l12_validation_5seed/paired_final_deltas.csv
ADDED
|
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
seed,condition,final_val,best_static_condition,best_static_final_val,delta_vs_best_static
|
| 2 |
+
1,wikitext103_formula_l12,4.062299974262714,static_dropout_0.1,4.080718021839857,-0.018418047577142715
|
| 3 |
+
1,wikitext103_probe_blend,4.07382857054472,static_dropout_0.1,4.080718021839857,-0.006889451295137405
|
| 4 |
+
1,wikitext103_low_decay,4.085419669747353,static_dropout_0.1,4.080718021839857,0.004701647907495499
|
| 5 |
+
1,static_dropout_0.1,4.080718021839857,static_dropout_0.1,4.080718021839857,0.0
|
| 6 |
+
1,static_dropout_0.08,4.089278034865856,static_dropout_0.1,4.080718021839857,0.00856001302599907
|
| 7 |
+
1,static_dropout_0.06,4.111171845346689,static_dropout_0.1,4.080718021839857,0.030453823506832123
|
| 8 |
+
1,static_dropout_0.14,4.108215663582087,static_dropout_0.1,4.080718021839857,0.02749764174222946
|
| 9 |
+
1,static_dropout_0.18,4.110849380493164,static_dropout_0.1,4.080718021839857,0.03013135865330696
|
| 10 |
+
1,static_dropout_0.2,4.116236750036478,static_dropout_0.1,4.080718021839857,0.03551872819662094
|
| 11 |
+
1,static_dropout_0.04,4.103101845830679,static_dropout_0.1,4.080718021839857,0.02238382399082184
|
| 12 |
+
1,static_dropout_0.02,4.137094188481569,static_dropout_0.1,4.080718021839857,0.05637616664171219
|
| 13 |
+
1,static_dropout_0,4.159992527216673,static_dropout_0.1,4.080718021839857,0.0792745053768158
|
| 14 |
+
1,static_dropout_0.26,4.155716687440872,static_dropout_0.1,4.080718021839857,0.07499866560101509
|
| 15 |
+
1,static_dropout_0.3,4.180164486169815,static_dropout_0.1,4.080718021839857,0.09944646432995796
|
| 16 |
+
2,wikitext103_formula_l12,4.112260434776545,static_dropout_0.06,4.130373623222113,-0.018113188445568085
|
| 17 |
+
2,wikitext103_probe_blend,4.111278761178255,static_dropout_0.06,4.130373623222113,-0.019094862043857574
|
| 18 |
+
2,wikitext103_low_decay,4.129110310226679,static_dropout_0.06,4.130373623222113,-0.0012633129954338074
|
| 19 |
+
2,static_dropout_0.1,4.131969079375267,static_dropout_0.06,4.130373623222113,0.0015954561531543732
|
| 20 |
+
2,static_dropout_0.08,4.137437805533409,static_dropout_0.06,4.130373623222113,0.007064182311296463
|
| 21 |
+
2,static_dropout_0.06,4.130373623222113,static_dropout_0.06,4.130373623222113,0.0
|
| 22 |
+
2,static_dropout_0.14,4.147622540593147,static_dropout_0.06,4.130373623222113,0.017248917371034622
|
| 23 |
+
2,static_dropout_0.18,4.147120479494333,static_dropout_0.06,4.130373623222113,0.01674685627222061
|
| 24 |
+
2,static_dropout_0.2,4.163300380110741,static_dropout_0.06,4.130373623222113,0.032926756888628006
|
| 25 |
+
2,static_dropout_0.04,4.164802756160498,static_dropout_0.06,4.130373623222113,0.03442913293838501
|
| 26 |
+
2,static_dropout_0.02,4.174610733985901,static_dropout_0.06,4.130373623222113,0.04423711076378822
|
| 27 |
+
2,static_dropout_0,4.202965669333935,static_dropout_0.06,4.130373623222113,0.07259204611182213
|
| 28 |
+
2,static_dropout_0.26,4.1961489245295525,static_dropout_0.06,4.130373623222113,0.0657753013074398
|
| 29 |
+
2,static_dropout_0.3,4.215538114309311,static_dropout_0.06,4.130373623222113,0.08516449108719826
|
| 30 |
+
3,wikitext103_formula_l12,4.0763493329286575,static_dropout_0.08,4.1035647466778755,-0.027215413749217987
|
| 31 |
+
3,wikitext103_probe_blend,4.093361176550388,static_dropout_0.08,4.1035647466778755,-0.010203570127487183
|
| 32 |
+
3,wikitext103_low_decay,4.100566737353802,static_dropout_0.08,4.1035647466778755,-0.0029980093240737915
|
| 33 |
+
3,static_dropout_0.1,4.111502721905708,static_dropout_0.08,4.1035647466778755,0.007937975227832794
|
| 34 |
+
3,static_dropout_0.08,4.1035647466778755,static_dropout_0.08,4.1035647466778755,0.0
|
| 35 |
+
3,static_dropout_0.06,4.112732540816069,static_dropout_0.08,4.1035647466778755,0.00916779413819313
|
| 36 |
+
3,static_dropout_0.14,4.123979520052671,static_dropout_0.08,4.1035647466778755,0.020414773374795914
|
| 37 |
+
3,static_dropout_0.18,4.128544177860022,static_dropout_0.08,4.1035647466778755,0.024979431182146072
|
| 38 |
+
3,static_dropout_0.2,4.136738710105419,static_dropout_0.08,4.1035647466778755,0.03317396342754364
|
| 39 |
+
3,static_dropout_0.04,4.124624028801918,static_dropout_0.08,4.1035647466778755,0.02105928212404251
|
| 40 |
+
3,static_dropout_0.02,4.144321486353874,static_dropout_0.08,4.1035647466778755,0.04075673967599869
|
| 41 |
+
3,static_dropout_0,4.175786443054676,static_dropout_0.08,4.1035647466778755,0.07222169637680054
|
| 42 |
+
3,static_dropout_0.26,4.179637394845486,static_dropout_0.08,4.1035647466778755,0.07607264816761017
|
| 43 |
+
3,static_dropout_0.3,4.192550577223301,static_dropout_0.08,4.1035647466778755,0.08898583054542542
|
| 44 |
+
4,wikitext103_formula_l12,4.0844879895448685,static_dropout_0.1,4.109558492898941,-0.02507050335407257
|
| 45 |
+
4,wikitext103_probe_blend,4.09542103484273,static_dropout_0.1,4.109558492898941,-0.014137458056211472
|
| 46 |
+
4,wikitext103_low_decay,4.092821758240461,static_dropout_0.1,4.109558492898941,-0.01673673465847969
|
| 47 |
+
4,static_dropout_0.1,4.109558492898941,static_dropout_0.1,4.109558492898941,0.0
|
| 48 |
+
4,static_dropout_0.08,4.122252244502306,static_dropout_0.1,4.109558492898941,0.012693751603364944
|
| 49 |
+
4,static_dropout_0.06,4.118809700012207,static_dropout_0.1,4.109558492898941,0.009251207113265991
|
| 50 |
+
4,static_dropout_0.14,4.111692663282156,static_dropout_0.1,4.109558492898941,0.0021341703832149506
|
| 51 |
+
4,static_dropout_0.18,4.132994774729013,static_dropout_0.1,4.109558492898941,0.023436281830072403
|
| 52 |
+
4,static_dropout_0.2,4.13876885920763,static_dropout_0.1,4.109558492898941,0.029210366308689117
|
| 53 |
+
4,static_dropout_0.04,4.131225533783436,static_dropout_0.1,4.109558492898941,0.02166704088449478
|
| 54 |
+
4,static_dropout_0.02,4.1386831142008305,static_dropout_0.1,4.109558492898941,0.02912462130188942
|
| 55 |
+
4,static_dropout_0,4.185262691229582,static_dropout_0.1,4.109558492898941,0.07570419833064079
|
| 56 |
+
4,static_dropout_0.26,4.178157042711973,static_dropout_0.1,4.109558492898941,0.06859854981303215
|
| 57 |
+
4,static_dropout_0.3,4.200709246098995,static_dropout_0.1,4.109558492898941,0.09115075320005417
|
| 58 |
+
5,wikitext103_formula_l12,4.068591594696045,static_dropout_0.08,4.105628222227097,-0.037036627531051636
|
| 59 |
+
5,wikitext103_probe_blend,4.106558587402105,static_dropout_0.08,4.105628222227097,0.0009303651750087738
|
| 60 |
+
5,wikitext103_low_decay,4.102147191762924,static_dropout_0.08,4.105628222227097,-0.0034810304641723633
|
| 61 |
+
5,static_dropout_0.1,4.118561428040266,static_dropout_0.08,4.105628222227097,0.01293320581316948
|
| 62 |
+
5,static_dropout_0.08,4.105628222227097,static_dropout_0.08,4.105628222227097,0.0
|
| 63 |
+
5,static_dropout_0.06,4.125324193388224,static_dropout_0.08,4.105628222227097,0.01969597116112709
|
| 64 |
+
5,static_dropout_0.14,4.1191585548222065,static_dropout_0.08,4.105628222227097,0.01353033259510994
|
| 65 |
+
5,static_dropout_0.18,4.132549241185188,static_dropout_0.08,4.105628222227097,0.026921018958091736
|
| 66 |
+
5,static_dropout_0.2,4.141841117292643,static_dropout_0.08,4.105628222227097,0.036212895065546036
|
| 67 |
+
5,static_dropout_0.04,4.141878347843885,static_dropout_0.08,4.105628222227097,0.036250125616788864
|
| 68 |
+
5,static_dropout_0.02,4.134641747921705,static_dropout_0.08,4.105628222227097,0.02901352569460869
|
| 69 |
+
5,static_dropout_0,4.193427540361881,static_dropout_0.08,4.105628222227097,0.0877993181347847
|
| 70 |
+
5,static_dropout_0.26,4.18209795281291,static_dropout_0.08,4.105628222227097,0.07646973058581352
|
| 71 |
+
5,static_dropout_0.3,4.184087038040161,static_dropout_0.08,4.105628222227097,0.07845881581306458
|
runs/wikitext103_streaming_report/l12_validation_5seed/stage_summary.csv
ADDED
|
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
condition,stage,token_limit,dropout,n,mean_val,std_val,mean_train,std_train,mean_gap,std_gap
|
| 2 |
+
wikitext103_formula_l12,0,250000,0.3,5,5.21483291387558,0.018060268226475495,4.213146224617958,0.013218263380840874,1.0016866892576217,0.027238437205160136
|
| 3 |
+
wikitext103_formula_l12,1,500000,0.26,5,4.80534844994545,0.027829930328897787,3.918210099637508,0.011987444359375697,0.8871383503079414,0.02626272167699895
|
| 4 |
+
wikitext103_formula_l12,2,1000000,0.18,5,4.49377838075161,0.014674108647494532,3.8283171117305757,0.017535628899982475,0.6654612690210342,0.02707089996550965
|
| 5 |
+
wikitext103_formula_l12,3,2000000,0.09,5,4.260725708305836,0.018098739272622868,3.8088956013321877,0.016376701234484724,0.45183010697364806,0.01960955973506382
|
| 6 |
+
wikitext103_formula_l12,4,4000000,0.02,5,4.080797865241766,0.019458062276034267,3.7991417959332465,0.028273745467366926,0.2816560693085194,0.038849358858741
|
| 7 |
+
wikitext103_probe_blend,0,250000,0.19,5,5.165890334546566,0.017075109484847858,4.020059056580067,0.016213485710936643,1.1458312779664994,0.02016518864116932
|
| 8 |
+
wikitext103_probe_blend,1,500000,0.14,5,4.7872447416186334,0.0268838649061173,3.684571087360382,0.02397149156090502,1.1026736542582511,0.021535214146850536
|
| 9 |
+
wikitext103_probe_blend,2,1000000,0.09,5,4.49399798810482,0.015864095933133645,3.6495467707514764,0.019490464893964263,0.8444512173533439,0.03475725994195922
|
| 10 |
+
wikitext103_probe_blend,3,2000000,0.04,5,4.274319227784872,0.0173783813582558,3.720025636255741,0.012889615922098394,0.5542935915291309,0.012746502538982045
|
| 11 |
+
wikitext103_probe_blend,4,4000000,0.01,5,4.096089626103639,0.014523730760550201,3.7674024373292925,0.027214408565503653,0.3286871887743473,0.038218075738052845
|
| 12 |
+
wikitext103_low_decay,0,250000,0.14,5,5.163547059893608,0.02195330299872749,3.9050611212849615,0.01328398150034886,1.2584859386086464,0.025343680837745006
|
| 13 |
+
wikitext103_low_decay,1,500000,0.14,5,4.800087235867977,0.01984656043518159,3.6649743929505347,0.023120354646606448,1.1351128429174424,0.01298098541252223
|
| 14 |
+
wikitext103_low_decay,2,1000000,0.1,5,4.50133255571127,0.01578619433972172,3.6606912568211554,0.023899445408052746,0.8406412988901139,0.03886870429434243
|
| 15 |
+
wikitext103_low_decay,3,2000000,0.06,5,4.273568108677864,0.022770462958788148,3.7474397838115694,0.003830485352807811,0.5261283248662949,0.02051193157067845
|
| 16 |
+
wikitext103_low_decay,4,4000000,0.02,5,4.102013133466244,0.016554390946653275,3.776866267621517,0.02647416884246062,0.3251468658447266,0.03314100249583806
|
| 17 |
+
static_dropout_0.1,0,250000,0.1,5,5.169893845915794,0.023739040366881205,3.8219083666801454,0.019248328397722874,1.3479854792356492,0.03169483270954973
|
| 18 |
+
static_dropout_0.1,1,500000,0.1,5,4.83315060287714,0.025840187319246186,3.5636927232146265,0.024862277351108786,1.2694578796625138,0.01835928668354016
|
| 19 |
+
static_dropout_0.1,2,1000000,0.1,5,4.518611620366573,0.01351751181040844,3.652407945692539,0.022251919818600654,0.8662036746740341,0.033766642849086155
|
| 20 |
+
static_dropout_0.1,3,2000000,0.1,5,4.285703422129155,0.018994111876547076,3.7808884128928186,0.01766528157534878,0.5048150092363357,0.028072111875115907
|
| 21 |
+
static_dropout_0.1,4,4000000,0.1,5,4.1104619488120075,0.018806640195540816,3.841722105443478,0.019884552697005082,0.2687398433685303,0.03219305563944893
|
| 22 |
+
static_dropout_0.08,0,250000,0.08,5,5.189354091882706,0.016114118388750917,3.761947725713253,0.017361993699706244,1.4274063661694527,0.012218707428925264
|
| 23 |
+
static_dropout_0.08,1,500000,0.08,5,4.857585413753986,0.023903772104159438,3.50356438010931,0.024856947856236954,1.3540210336446763,0.0310369653543579
|
| 24 |
+
static_dropout_0.08,2,1000000,0.08,5,4.532606348395348,0.011746366682868653,3.5999757796525955,0.009078845784580064,0.9326305687427521,0.019403758262432406
|
| 25 |
+
static_dropout_0.08,3,2000000,0.08,5,4.292461846768856,0.016558424281915483,3.7518652200698854,0.019831550751639896,0.5405966266989708,0.034040566494281235
|
| 26 |
+
static_dropout_0.08,4,4000000,0.08,5,4.111632210761309,0.01857117977689519,3.8268220275640488,0.021203534366681782,0.2848101831972599,0.03679859980086126
|
| 27 |
+
static_dropout_0.06,0,250000,0.06,5,5.215402702987194,0.01734079554543704,3.712760145962238,0.018437630763482243,1.5026425570249557,0.03154944432034942
|
| 28 |
+
static_dropout_0.06,1,500000,0.06,5,4.894671627879143,0.021331675814859084,3.4394222095608713,0.013418585935883633,1.4552494183182716,0.01946001354297248
|
| 29 |
+
static_dropout_0.06,2,1000000,0.06,5,4.557375983893872,0.012636266859277006,3.555374290049076,0.018874416206200166,1.0020016938447953,0.030341255269329737
|
| 30 |
+
static_dropout_0.06,3,2000000,0.06,5,4.305900506675243,0.018234543669842573,3.7317687764763834,0.017157479537739013,0.5741317301988602,0.023633182905307998
|
| 31 |
+
static_dropout_0.06,4,4000000,0.06,5,4.119682380557061,0.008179605355252681,3.806561966240406,0.022924307415498075,0.3131204143166542,0.028667268842643943
|
| 32 |
+
static_dropout_0.14,0,250000,0.14,5,5.1635470792651175,0.021953339409384615,3.9050611019134522,0.013283976741307785,1.2584859773516655,0.025343683677925957
|
| 33 |
+
static_dropout_0.14,1,500000,0.14,5,4.800087215006352,0.019846558873074828,3.6649743527173997,0.023120394246437293,1.135112862288952,0.012980966896977213
|
| 34 |
+
static_dropout_0.14,2,1000000,0.14,5,4.500055414438248,0.014819407090300617,3.716250878572464,0.02407671653248301,0.7838045358657837,0.038101035301189086
|
| 35 |
+
static_dropout_0.14,3,2000000,0.14,5,4.281563547253609,0.02138858846966249,3.828673002123833,0.004937581309037719,0.452890545129776,0.01755457638984346
|
| 36 |
+
static_dropout_0.14,4,4000000,0.14,5,4.122133788466454,0.015529540565590345,3.8673695802688597,0.025363134860510252,0.2547642081975937,0.0302072467927019
|
| 37 |
+
static_dropout_0.18,0,250000,0.18,5,5.16162400841713,0.014987630048664583,3.9964362382888794,0.01862825818296394,1.16518777012825,0.02013674124988852
|
| 38 |
+
static_dropout_0.18,1,500000,0.18,5,4.794638857245445,0.02062163611688947,3.7571836963295935,0.024732309568709827,1.0374551609158515,0.02312552092553772
|
| 39 |
+
static_dropout_0.18,2,1000000,0.18,5,4.5023036181926726,0.018490102933264013,3.786629955470562,0.015769997442310384,0.7156736627221107,0.03180040950921459
|
| 40 |
+
static_dropout_0.18,3,2000000,0.18,5,4.288915157318115,0.019532110178580833,3.876796562969685,0.008357602777023191,0.4121185943484306,0.013904819825939841
|
| 41 |
+
static_dropout_0.18,4,4000000,0.18,5,4.130411610752344,0.013005735782293457,3.9014849200844766,0.018144303418082167,0.22892669066786767,0.02966442514236336
|
| 42 |
+
static_dropout_0.2,0,250000,0.2,5,5.170051643252373,0.014071496005973716,4.036272630095482,0.013668609797313208,1.1337790131568908,0.013715020568700555
|
| 43 |
+
static_dropout_0.2,1,500000,0.2,5,4.787301687896251,0.023601504941398347,3.7914333969354628,0.018987144953080883,0.9958682909607888,0.014572669955232898
|
| 44 |
+
static_dropout_0.2,2,1000000,0.2,5,4.506029562652111,0.02043622151331783,3.814779002964497,0.01587114098616348,0.6912505596876144,0.03147408021666257
|
| 45 |
+
static_dropout_0.2,3,2000000,0.2,5,4.294206875562668,0.016019909189912813,3.898232290148735,0.008893270752025355,0.3959745854139328,0.01268490098888804
|
| 46 |
+
static_dropout_0.2,4,4000000,0.2,5,4.1393771633505825,0.016742174044810716,3.9154964342713354,0.020116706552102667,0.22388072907924653,0.033233695598367155
|
| 47 |
+
static_dropout_0.04,0,250000,0.04,5,5.237822580337524,0.018592151453448197,3.6440532103180887,0.031242522590923742,1.593769370019436,0.03528512909650908
|
| 48 |
+
static_dropout_0.04,1,500000,0.04,5,4.957330641150475,0.02503345339488141,3.3515226930379867,0.029537749484816224,1.6058079481124878,0.0202359151612378
|
| 49 |
+
static_dropout_0.04,2,1000000,0.04,5,4.595940679311752,0.014619476464690098,3.502993068099022,0.017291674978225223,1.0929476112127303,0.028916946275823
|
| 50 |
+
static_dropout_0.04,3,2000000,0.04,5,4.3262467920780185,0.016533384756725625,3.703781445324421,0.01970997472996272,0.6224653467535972,0.02577206799431021
|
| 51 |
+
static_dropout_0.04,4,4000000,0.04,5,4.1331265024840835,0.022679096058077455,3.7978432700037956,0.018504253509089505,0.3352832324802876,0.036956450676459744
|
| 52 |
+
static_dropout_0.02,0,250000,0.02,5,5.274957031011581,0.025475383539799955,3.572465108335018,0.027061999609254494,1.7024919226765634,0.03656967535619302
|
| 53 |
+
static_dropout_0.02,1,500000,0.02,5,5.045114178955555,0.016894569246234076,3.261230443418026,0.014168838694048941,1.783883735537529,0.019840865900389654
|
| 54 |
+
static_dropout_0.02,2,1000000,0.02,5,4.655830132961273,0.01585338982667534,3.4323942199349404,0.018124036007212796,1.2234359130263328,0.024146051923035036
|
| 55 |
+
static_dropout_0.02,3,2000000,0.02,5,4.355083760619164,0.026489436313039863,3.6672953754663467,0.010966930251295475,0.6877883851528168,0.026106538952549097
|
| 56 |
+
static_dropout_0.02,4,4000000,0.02,5,4.145870254188776,0.016455755256661365,3.775918734073639,0.03360785862475549,0.3699515201151371,0.0368559849747316
|
| 57 |
+
static_dropout_0,0,250000,0.0,5,5.340271946787834,0.02699193219713777,3.5230478435754775,0.03634617773629946,1.8172241032123566,0.05178790589031804
|
| 58 |
+
static_dropout_0,1,500000,0.0,5,5.174065832793713,0.025168511150111032,3.1505795806646346,0.025699619025332174,2.0234862521290777,0.0236535174506497
|
| 59 |
+
static_dropout_0,2,1000000,0.0,5,4.766069588065148,0.03242686059194632,3.3657630145549775,0.025194332688414405,1.40030657351017,0.052237334934698405
|
| 60 |
+
static_dropout_0,3,2000000,0.0,5,4.417352265119552,0.01876483626309062,3.6409011498093604,0.014124132971310668,0.7764511153101921,0.024211298020746064
|
| 61 |
+
static_dropout_0,4,4000000,0.0,5,4.183486974239349,0.01652781602088682,3.7750062167644503,0.025453964920340872,0.4084807574748993,0.03358628463770181
|
| 62 |
+
static_dropout_0.26,0,250000,0.26,5,5.194025552272796,0.016128282325428966,4.149615630507469,0.024675006962846267,1.0444099217653275,0.022196467420881313
|
| 63 |
+
static_dropout_0.26,1,500000,0.26,5,4.8081169500947,0.021576250073037834,3.905269515514374,0.02376505939443177,0.9028474345803261,0.018700407757224247
|
| 64 |
+
static_dropout_0.26,2,1000000,0.26,5,4.52620010226965,0.010149133513914643,3.9071271896362303,0.02774032700763299,0.619072912633419,0.035195638721522884
|
| 65 |
+
static_dropout_0.26,3,2000000,0.26,5,4.32477528527379,0.017871972287011016,3.9654731526970863,0.018160558354021094,0.359302132576704,0.018728321745075875
|
| 66 |
+
static_dropout_0.26,4,4000000,0.26,5,4.178351600468159,0.01453281035529064,3.9775424867868425,0.021977989750098077,0.20080911368131638,0.03295020359909027
|
| 67 |
+
static_dropout_0.3,0,250000,0.3,5,5.214832927286625,0.018060234410376495,4.213146212697029,0.013218280901902867,1.0016867145895958,0.02723842931339789
|
| 68 |
+
static_dropout_0.3,1,500000,0.3,5,4.82416096329689,0.029627048936626245,3.976513123512268,0.010796527385978102,0.8476478397846222,0.026893359245285242
|
| 69 |
+
static_dropout_0.3,2,1000000,0.3,5,4.546238152682781,0.012672630425565377,3.970817744731903,0.017137273586986676,0.5754204079508781,0.02420747470801454
|
| 70 |
+
static_dropout_0.3,3,2000000,0.3,5,4.34667577445507,0.015854348211857338,4.009651578962803,0.014639843511797184,0.3370241954922676,0.019075525784314876
|
| 71 |
+
static_dropout_0.3,4,4000000,0.3,5,4.194609892368317,0.014131822174274241,4.012722708284855,0.024177126067976527,0.18188718408346177,0.031100561007876955
|