Upload docs/EXT32K_RESULTS_2026-09-09.md with huggingface_hub
Browse files
docs/EXT32K_RESULTS_2026-09-09.md
ADDED
|
@@ -0,0 +1,56 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# 16k -> 32k YaRN extension on b4 -- 2026-09-09
|
| 2 |
+
|
| 3 |
+
Parent: l2 (lenbucket_l2_blumodelcp_run2_step_400.ckpt), the best 16k arm.
|
| 4 |
+
Data: b4_32768_65536 -- min doc 32,768 == the window, so EVERY training
|
| 5 |
+
sample is one contiguous document and a 32k-spanning dependency is
|
| 6 |
+
possible. b3 (min 16,384) would have straddled an EOT in every window.
|
| 7 |
+
YaRN: SCALE=2 ORIG_MAXPOS=16384 -> 32768 exactly.
|
| 8 |
+
Train: 750 steps, global batch 524,288 (B=1 x accum 16), WSD stable 550
|
| 9 |
+
+ 100-step decay, 2x H200 cp=2, peak 108.8 GB of 143 GB.
|
| 10 |
+
Final loss 2.570224. 393 M tokens of 586 M available (no repetition).
|
| 11 |
+
|
| 12 |
+
## niah_single, N=100, ONE haystack (b2 val)
|
| 13 |
+
|
| 14 |
+
| context | l2 (16k model, 16k cache) | ext32k (32k model, 32k cache) | delta |
|
| 15 |
+
|---|---|---|---|
|
| 16 |
+
| 4,096 | 80.0 | 75.0 | -5.0 |
|
| 17 |
+
| 8,192 | 62.0 | 60.0 | -2.0 |
|
| 18 |
+
| 16,360 | 35.0 | 45.0 | +10.0 |
|
| 19 |
+
| 32,744 | cannot run | **27.0** | new capability |
|
| 20 |
+
|
| 21 |
+
95% Wilson at N=100 is +-9 to 10 pp, so the combined bar is +-11 to 13 pp:
|
| 22 |
+
NONE of the three deltas is individually resolvable. The 32,744 row is not a
|
| 23 |
+
delta -- the parent's RoPE cache is sized 16,384 and cannot process 32k at all.
|
| 24 |
+
|
| 25 |
+
## KNOWN BUG in this run's raw summary.csv files
|
| 26 |
+
|
| 27 |
+
ext32k_b4.sh line 128 passed YARN_ORIG=16384 YARN=2 to BOTH models, so the
|
| 28 |
+
l2 column in ext32k_len_*/summary.csv is a MISMATCHED readout -- l2 was
|
| 29 |
+
trained at scale 4 / orig 4096. Measured damage: 4k -4.0, 8k -26.0, 16k -15.0.
|
| 30 |
+
The l2 numbers in the table above come from the matched H100 decision eval.
|
| 31 |
+
DO NOT use the l2 column in the raw CSVs.
|
| 32 |
+
|
| 33 |
+
### PROMPT_LEN=4096 (raw -- l2 column mismatched, see above)
|
| 34 |
+
```
|
| 35 |
+
task,ext32k,l2
|
| 36 |
+
niah_single,75.0,76.0
|
| 37 |
+
```
|
| 38 |
+
|
| 39 |
+
### PROMPT_LEN=8192 (raw -- l2 column mismatched, see above)
|
| 40 |
+
```
|
| 41 |
+
task,ext32k,l2
|
| 42 |
+
niah_single,60.0,36.0
|
| 43 |
+
```
|
| 44 |
+
|
| 45 |
+
### PROMPT_LEN=16360 (raw -- l2 column mismatched, see above)
|
| 46 |
+
```
|
| 47 |
+
task,ext32k,l2
|
| 48 |
+
niah_single,45.0,20.0
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
### PROMPT_LEN=32744 (raw -- l2 column mismatched, see above)
|
| 52 |
+
```
|
| 53 |
+
task,ext32k,l2
|
| 54 |
+
niah_single,27.0,11.0
|
| 55 |
+
```
|
| 56 |
+
|