Upload rl_training_log.txt with huggingface_hub
Browse files- rl_training_log.txt +493 -0
rl_training_log.txt
ADDED
|
@@ -0,0 +1,493 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 0 |
...build/ar_sft_shuf.parquet: 100%|ββββββββββ| 15.7MB / 15.7MB [A[A
|
|
|
|
|
|
|
|
|
|
| 1 |
...build/av_sft_shuf.parquet: 100%|ββββββββββ| 15.8MB / 15.8MB [A[A[A
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
...ata/build/rl_shuf.parquet: 100%|ββββββββββ| 30.1MB / 30.1MB [A[A[A[A
|
|
|
|
|
|
|
| 3 |
...build/ar_sft_shuf.parquet: 100%|ββββββββββ| 15.7MB / 15.7MB [A[A
|
|
|
|
|
|
|
|
|
|
| 4 |
...build/av_sft_shuf.parquet: 100%|ββββββββββ| 15.8MB / 15.8MB [A[A[A
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
...ata/build/rl_shuf.parquet: 100%|ββββββββββ| 30.1MB / 30.1MB [A[A[A[A
|
|
|
|
|
|
|
| 6 |
...build/ar_sft_shuf.parquet: 100%|ββββββββββ| 15.7MB / 15.7MB
|
|
|
|
| 7 |
...build/av_sft_shuf.parquet: 100%|ββββββββββ| 15.8MB / 15.8MB
|
|
|
|
| 8 |
...ata/build/rl_shuf.parquet: 100%|ββββββββββ| 30.1MB / 30.1MB
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 9 |
...adapter_model.safetensors: 18%|ββ | 88.0MB / 491MB [A[A
|
|
|
|
|
|
|
|
|
|
| 10 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB [A[A[A
|
|
|
|
|
|
|
| 11 |
...adapter_model.safetensors: 18%|ββ | 88.0MB / 491MB [A[A
|
|
|
|
|
|
|
|
|
|
| 12 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB [A[A[A
|
|
|
|
|
|
|
| 13 |
...adapter_model.safetensors: 34%|ββββ | 168MB / 491MB [A[A
|
|
|
|
|
|
|
|
|
|
| 14 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB [A[A[A
|
|
|
|
|
|
|
| 15 |
...adapter_model.safetensors: 52%|ββββββ | 256MB / 491MB [A[A
|
|
|
|
|
|
|
|
|
|
| 16 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB [A[A[A
|
|
|
|
|
|
|
| 17 |
...adapter_model.safetensors: 68%|βββββββ | 336MB / 491MB [A[A
|
|
|
|
|
|
|
|
|
|
| 18 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB [A[A[A
|
|
|
|
|
|
|
| 19 |
...adapter_model.safetensors: 86%|βββββββββ | 424MB / 491MB [A[A
|
|
|
|
|
|
|
|
|
|
| 20 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB [A[A[A
|
|
|
|
|
|
|
| 21 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB [A[A
|
|
|
|
|
|
|
|
|
|
| 22 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB [A[A[A
|
|
|
|
|
|
|
| 23 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB
|
|
|
|
| 24 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 25 |
...ra_value_head.safetensors: 39%|ββββ | 79.9MB / 204MB [A[A
|
|
|
|
|
|
|
|
|
|
| 26 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB [A[A[A
|
|
|
|
|
|
|
| 27 |
...ra_value_head.safetensors: 39%|ββββ | 79.9MB / 204MB [A[A
|
|
|
|
|
|
|
|
|
|
| 28 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB [A[A[A
|
|
|
|
|
|
|
| 29 |
...ra_value_head.safetensors: 82%|βββββββββ | 168MB / 204MB [A[A
|
|
|
|
|
|
|
|
|
|
| 30 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB [A[A[A
|
|
|
|
|
|
|
| 31 |
...ra_value_head.safetensors: 100%|ββββββββββ| 204MB / 204MB [A[A
|
|
|
|
|
|
|
|
|
|
| 32 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB [A[A[A
|
|
|
|
|
|
|
| 33 |
...ra_value_head.safetensors: 100%|ββββββββββ| 204MB / 204MB
|
|
|
|
| 34 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 35 |
...adapter_model.safetensors: 0%| | 10.5kB / 491MB [A[A
|
|
|
|
|
|
|
|
|
|
| 36 |
...adapter_model.safetensors: 16%|ββ | 79.9MB / 491MB [A[A[A
|
|
|
|
|
|
|
| 37 |
...adapter_model.safetensors: 0%| | 10.5kB / 491MB [A[A
|
|
|
|
|
|
|
|
|
|
| 38 |
...adapter_model.safetensors: 16%|ββ | 79.9MB / 491MB [A[A[A
|
|
|
|
|
|
|
| 39 |
...adapter_model.safetensors: 0%| | 10.5kB / 491MB [A[A
|
|
|
|
|
|
|
|
|
|
| 40 |
...adapter_model.safetensors: 34%|ββββ | 168MB / 491MB [A[A[A
|
|
|
|
|
|
|
| 41 |
...adapter_model.safetensors: 0%| | 10.5kB / 491MB [A[A
|
|
|
|
|
|
|
|
|
|
| 42 |
...adapter_model.safetensors: 52%|ββββββ | 256MB / 491MB [A[A[A
|
|
|
|
|
|
|
|
|
|
|
|
|
| 43 |
...ra_value_head.safetensors: 43%|βββββ | 87.9MB / 204MB [A[A[A[A
|
|
|
|
|
|
|
| 44 |
...adapter_model.safetensors: 21%|ββ | 104MB / 491MB [A[A
|
|
|
|
|
|
|
|
|
|
| 45 |
...adapter_model.safetensors: 70%|βββββββ | 344MB / 491MB [A[A[A
|
|
|
|
|
|
|
|
|
|
|
|
|
| 46 |
...ra_value_head.safetensors: 43%|βββββ | 87.9MB / 204MB [A[A[A[A
|
|
|
|
|
|
|
| 47 |
...adapter_model.safetensors: 37%|ββββ | 184MB / 491MB [A[A
|
|
|
|
|
|
|
|
|
|
| 48 |
...adapter_model.safetensors: 86%|βββββββββ | 424MB / 491MB [A[A[A
|
|
|
|
|
|
|
|
|
|
|
|
|
| 49 |
...ra_value_head.safetensors: 86%|βββββββββ | 176MB / 204MB [A[A[A[A
|
|
|
|
|
|
|
| 50 |
...adapter_model.safetensors: 54%|ββββββ | 264MB / 491MB [A[A
|
|
|
|
|
|
|
|
|
|
| 51 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB [A[A[A
|
|
|
|
|
|
|
|
|
|
|
|
|
| 52 |
...ra_value_head.safetensors: 100%|ββββββββββ| 204MB / 204MB [A[A[A[A
|
|
|
|
|
|
|
| 53 |
...adapter_model.safetensors: 70%|βββββββ | 344MB / 491MB [A[A
|
|
|
|
|
|
|
|
|
|
| 54 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB [A[A[A
|
|
|
|
|
|
|
|
|
|
|
|
|
| 55 |
...ra_value_head.safetensors: 100%|ββββββββββ| 204MB / 204MB [A[A[A[A
|
|
|
|
|
|
|
| 56 |
...adapter_model.safetensors: 88%|βββββββββ | 432MB / 491MB [A[A
|
|
|
|
|
|
|
|
|
|
| 57 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB [A[A[A
|
|
|
|
|
|
|
|
|
|
|
|
|
| 58 |
...ra_value_head.safetensors: 100%|ββββββββββ| 204MB / 204MB [A[A[A[A
|
|
|
|
|
|
|
| 59 |
...adapter_model.safetensors: 98%|ββββββββββ| 483MB / 491MB [A[A
|
|
|
|
|
|
|
|
|
|
| 60 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB [A[A[A
|
|
|
|
|
|
|
|
|
|
|
|
|
| 61 |
...ra_value_head.safetensors: 100%|ββββββββββ| 204MB / 204MB [A[A[A[A
|
|
|
|
|
|
|
|
|
|
| 62 |
...adapter_model.safetensors: 99%|ββββββββββ| 485MB / 491MB [A[A
|
|
|
|
|
|
|
|
|
|
| 63 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB [A[A[A
|
|
|
|
|
|
|
|
|
|
|
|
|
| 64 |
...ra_value_head.safetensors: 100%|ββββββββββ| 204MB / 204MB [A[A[A[A
|
|
|
|
|
|
|
|
|
|
| 65 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB [A[A
|
|
|
|
|
|
|
|
|
|
| 66 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB [A[A[A
|
|
|
|
|
|
|
|
|
|
|
|
|
| 67 |
...ra_value_head.safetensors: 100%|ββββββββββ| 204MB / 204MB [A[A[A[A
|
|
|
|
|
|
|
|
|
|
| 68 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB [A[A
|
|
|
|
|
|
|
|
|
|
| 69 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB [A[A[A
|
|
|
|
|
|
|
|
|
|
|
|
|
| 70 |
...ra_value_head.safetensors: 100%|ββββββββββ| 204MB / 204MB [A[A[A[A
|
|
|
|
|
|
|
|
|
|
| 71 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB
|
|
|
|
| 72 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB
|
|
|
|
| 73 |
...ra_value_head.safetensors: 100%|ββββββββββ| 204MB / 204MB
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
|
| 2 |
+
=================== STAGE 1/8: download slim dataset (syvb/nla-warmstart-explanations-finefineweb-sonnet46) ===================
|
| 3 |
+
2026-06-22T06:20:17Z
|
| 4 |
+
(skip) /workspace/nla_poc/data/slim.parquet exists
|
| 5 |
+
|
| 6 |
+
=================== STAGE 2/8: regenerate activations ===================
|
| 7 |
+
2026-06-22T06:20:17Z
|
| 8 |
+
(skip) /workspace/nla_poc/data/base_with_activations.parquet exists
|
| 9 |
+
|
| 10 |
+
=================== STAGE 3/8: synthesize base sidecar ===================
|
| 11 |
+
2026-06-22T06:20:17Z
|
| 12 |
+
(skip) sidecar exists
|
| 13 |
+
|
| 14 |
+
=================== STAGE 4/8: document-level split 25/25/50 ===================
|
| 15 |
+
2026-06-22T06:20:17Z
|
| 16 |
+
(skip) split exists
|
| 17 |
+
|
| 18 |
+
=================== STAGE 5/8: build training parquets (av_sft, ar_sft, rl) ===================
|
| 19 |
+
2026-06-22T06:20:17Z
|
| 20 |
+
row counts:
|
| 21 |
+
av_sft_shuf: 2087 rows
|
| 22 |
+
ar_sft_shuf: 2090 rows
|
| 23 |
+
rl_shuf: 4220 rows
|
| 24 |
+
[push] /workspace/nla_poc/data/build -> syvb/nla-ordered-features-poc-data (dataset)
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
|
| 29 |
...build/ar_sft_shuf.parquet: 100%|ββββββββββ| 15.7MB / 15.7MB [A[A
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
|
| 33 |
...build/av_sft_shuf.parquet: 100%|ββββββββββ| 15.8MB / 15.8MB [A[A[A
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
|
| 37 |
+
|
| 38 |
...ata/build/rl_shuf.parquet: 100%|ββββββββββ| 30.1MB / 30.1MB [A[A[A[A
|
| 39 |
+
|
| 40 |
+
|
| 41 |
...build/ar_sft_shuf.parquet: 100%|ββββββββββ| 15.7MB / 15.7MB [A[A
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
|
| 45 |
...build/av_sft_shuf.parquet: 100%|ββββββββββ| 15.8MB / 15.8MB [A[A[A
|
| 46 |
+
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
|
| 50 |
...ata/build/rl_shuf.parquet: 100%|ββββββββββ| 30.1MB / 30.1MB [A[A[A[A
|
| 51 |
+
|
| 52 |
+
|
| 53 |
...build/ar_sft_shuf.parquet: 100%|ββββββββββ| 15.7MB / 15.7MB
|
| 54 |
+
|
| 55 |
...build/av_sft_shuf.parquet: 100%|ββββββββββ| 15.8MB / 15.8MB
|
| 56 |
+
|
| 57 |
...ata/build/rl_shuf.parquet: 100%|ββββββββββ| 30.1MB / 30.1MB
|
| 58 |
+
No files have been modified since last commit. Skipping to prevent empty commit.
|
| 59 |
+
[push_ckpt] pushed /workspace/nla_poc/data/build -> syvb/nla-ordered-features-poc-data (dataset)
|
| 60 |
+
[push_ckpt] added syvb/nla-ordered-features-poc-data to collection
|
| 61 |
+
|
| 62 |
+
=================== STAGE 6/8: AV warm-start SFT (300 steps) ===================
|
| 63 |
+
2026-06-22T06:20:19Z
|
| 64 |
+
(skip) AV checkpoint exists
|
| 65 |
+
AV_CKPT=/workspace/nla_poc/ckpts/av_sft/iter_0000300
|
| 66 |
+
[push] /workspace/nla_poc/ckpts/av_sft/iter_0000300 -> syvb/nla-ordered-features-av-sft (model)
|
| 67 |
+
|
| 68 |
+
|
| 69 |
+
|
| 70 |
+
|
| 71 |
...adapter_model.safetensors: 18%|ββ | 88.0MB / 491MB [A[A
|
| 72 |
+
|
| 73 |
+
|
| 74 |
+
|
| 75 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB [A[A[A
|
| 76 |
+
|
| 77 |
+
|
| 78 |
...adapter_model.safetensors: 18%|ββ | 88.0MB / 491MB [A[A
|
| 79 |
+
|
| 80 |
+
|
| 81 |
+
|
| 82 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB [A[A[A
|
| 83 |
+
|
| 84 |
+
|
| 85 |
...adapter_model.safetensors: 34%|ββββ | 168MB / 491MB [A[A
|
| 86 |
+
|
| 87 |
+
|
| 88 |
+
|
| 89 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB [A[A[A
|
| 90 |
+
|
| 91 |
+
|
| 92 |
...adapter_model.safetensors: 52%|ββββββ | 256MB / 491MB [A[A
|
| 93 |
+
|
| 94 |
+
|
| 95 |
+
|
| 96 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB [A[A[A
|
| 97 |
+
|
| 98 |
+
|
| 99 |
...adapter_model.safetensors: 68%|βββββββ | 336MB / 491MB [A[A
|
| 100 |
+
|
| 101 |
+
|
| 102 |
+
|
| 103 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB [A[A[A
|
| 104 |
+
|
| 105 |
+
|
| 106 |
...adapter_model.safetensors: 86%|βββββββββ | 424MB / 491MB [A[A
|
| 107 |
+
|
| 108 |
+
|
| 109 |
+
|
| 110 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB [A[A[A
|
| 111 |
+
|
| 112 |
+
|
| 113 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB [A[A
|
| 114 |
+
|
| 115 |
+
|
| 116 |
+
|
| 117 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB [A[A[A
|
| 118 |
+
|
| 119 |
+
|
| 120 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB
|
| 121 |
+
|
| 122 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB
|
| 123 |
+
No files have been modified since last commit. Skipping to prevent empty commit.
|
| 124 |
+
[push_ckpt] pushed /workspace/nla_poc/ckpts/av_sft/iter_0000300 -> syvb/nla-ordered-features-av-sft (model)
|
| 125 |
+
[push_ckpt] added syvb/nla-ordered-features-av-sft to collection
|
| 126 |
+
|
| 127 |
+
=================== STAGE 7/8: AR warm-start SFT (300 steps) ===================
|
| 128 |
+
2026-06-22T06:20:22Z
|
| 129 |
+
(skip) AR checkpoint exists
|
| 130 |
+
AR_CKPT=/workspace/nla_poc/ckpts/ar_sft/iter_0000300
|
| 131 |
+
[push] /workspace/nla_poc/ckpts/ar_sft/iter_0000300 -> syvb/nla-ordered-features-ar-sft (model)
|
| 132 |
+
|
| 133 |
+
|
| 134 |
+
|
| 135 |
+
|
| 136 |
...ra_value_head.safetensors: 39%|ββββ | 79.9MB / 204MB [A[A
|
| 137 |
+
|
| 138 |
+
|
| 139 |
+
|
| 140 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB [A[A[A
|
| 141 |
+
|
| 142 |
+
|
| 143 |
...ra_value_head.safetensors: 39%|ββββ | 79.9MB / 204MB [A[A
|
| 144 |
+
|
| 145 |
+
|
| 146 |
+
|
| 147 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB [A[A[A
|
| 148 |
+
|
| 149 |
+
|
| 150 |
...ra_value_head.safetensors: 82%|βββββββββ | 168MB / 204MB [A[A
|
| 151 |
+
|
| 152 |
+
|
| 153 |
+
|
| 154 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB [A[A[A
|
| 155 |
+
|
| 156 |
+
|
| 157 |
...ra_value_head.safetensors: 100%|ββββββββββ| 204MB / 204MB [A[A
|
| 158 |
+
|
| 159 |
+
|
| 160 |
+
|
| 161 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB [A[A[A
|
| 162 |
+
|
| 163 |
+
|
| 164 |
...ra_value_head.safetensors: 100%|ββββββββββ| 204MB / 204MB
|
| 165 |
+
|
| 166 |
...er_0000300/tokenizer.json: 100%|ββββββββββ| 11.4MB / 11.4MB
|
| 167 |
+
No files have been modified since last commit. Skipping to prevent empty commit.
|
| 168 |
+
[push_ckpt] pushed /workspace/nla_poc/ckpts/ar_sft/iter_0000300 -> syvb/nla-ordered-features-ar-sft (model)
|
| 169 |
+
[push_ckpt] added syvb/nla-ordered-features-ar-sft to collection
|
| 170 |
+
|
| 171 |
+
=================== STAGE 8/8: RL ordered-features (post-hoc/per-group+maskK, 120 steps) ===================
|
| 172 |
+
2026-06-22T06:20:25Z
|
| 173 |
+
[cfg] inj_id=149705 mse_scale_f=64.0 d_model=4096
|
| 174 |
+
[actor] base=Qwen/Qwen3-8B + AV-LoRA=/workspace/nla_poc/ckpts/av_sft/iter_0000300 (quant=4bit, device_map=single)
|
| 175 |
+
`torch_dtype` is deprecated! Use `dtype` instead!
|
| 176 |
+
|
| 177 |
+
trainable params: 122,683,392 || all params: 8,436,102,144 || trainable%: 1.4543
|
| 178 |
+
[critic] AR-LoRA from /workspace/nla_poc/ckpts/ar_sft/iter_0000300: {'ar_num_layers': 25, 'lora_r': 128, 'lora_alpha': 16, 'quant': '4bit', 'target_modules': ['q_proj', 'k_proj', 'v_proj', 'o_proj'], 'final_norm_stripped': True}
|
| 179 |
+
|
| 180 |
+
[critic] truncated to 25 layers, value_head identity-init (weight norm = 64.000)
|
| 181 |
+
[critic] loaded 201 AR tensors (missing=276 backbone keys, unexpected=0)
|
| 182 |
+
[critic] CO-TRAINED, lr=5e-05, trainable=0.10B (backbone + value_head)
|
| 183 |
+
[critic] value_head shape=(4096, 4096)
|
| 184 |
+
[rollout] EOS-trim ids: [151643, 151645]
|
| 185 |
+
[data] loading /workspace/nla_poc/data/build/rl_shuf.parquet (max_rows=3000)
|
| 186 |
+
[data] 3000 rows
|
| 187 |
+
[fve] predict-the-mean baseline mse_nrm = 0.5603 (paper def; meannorm baseline = 0.6738)
|
| 188 |
+
[optim] using bitsandbytes AdamW8bit (bnb 0.49.2)
|
| 189 |
+
wandb: [wandb.login()] Loaded credentials for https://api.wandb.ai from WANDB_API_KEY.
|
| 190 |
+
wandb: Currently logged in as: syvb (octahedral-systems) to https://api.wandb.ai. Use `wandb login --relogin` to force relogin
|
| 191 |
+
wandb: Tracking run with wandb version 0.27.2
|
| 192 |
+
wandb: Run data is saved locally in /workspace/nanoNLA/wandb/run-20260622_062121-3h87i5bj
|
| 193 |
+
wandb: Run `wandb offline` to turn off syncing.
|
| 194 |
+
wandb: Syncing run rl_ordered_poc
|
| 195 |
+
wandb: βοΈ View project at https://wandb.ai/octahedral-systems/nla-qwen3-8b
|
| 196 |
+
wandb: π View run at https://wandb.ai/octahedral-systems/nla-qwen3-8b/runs/3h87i5bj
|
| 197 |
+
[eval] 20 doc-disjoint prompts loaded (rows past 3000, excluding 2827 training doc_ids)
|
| 198 |
+
[ordered-features] post-hoc/per-group + loss-mask-to-K: critic sees Uniform[1,10] features
|
| 199 |
+
step 0000 | loss 0.0004 | r -0.546 | FVE 2.6% | crit 0.5460 | kl 0.0000 | clip 0.00% | ext 69% | t 161s
|
| 200 |
+
[eval@0] reward -0.887 | FVE -23.2% | ext 85%
|
| 201 |
+
[eval@0 idx=0 r=-0.804] Date pattern "July [day]" requires day number like 1st/5th Specific recording date expected for British Academy explanation Narrative structure expects final chronological conclusion statement "existe
|
| 202 |
+
[eval@0 idx=7 r=-2.000] <extraction failed>
|
| 203 |
+
[eval@0 idx=14 r=-0.726] List item continuation: another event date expected Bullet-pointed program detail: recurring time/guest format Specific date pattern: "June [month] [date]" immediately follows Carolyn Brown biographic
|
| 204 |
+
step 0001 | loss 0.0004 | r -0.566 | FVE -1.0% | crit 0.5661 | kl 0.0108 | clip 0.00% | ext 83% | t 163s
|
| 205 |
+
step 0002 | loss 0.0003 | r -0.612 | FVE -9.2% | crit 0.6119 | kl 0.0349 | clip 0.00% | ext 97% | t 161s
|
| 206 |
+
step 0003 | loss 0.0003 | r -0.636 | FVE -13.5% | crit 0.6358 | kl 0.0455 | clip 0.00% | ext 98% | t 154s
|
| 207 |
+
step 0004 | loss 0.0015 | r -0.592 | FVE -5.6% | crit 0.5917 | kl 0.0928 | clip 0.00% | ext 98% | t 154s
|
| 208 |
+
step 0005 | loss 0.0019 | r -0.592 | FVE -5.6% | crit 0.5923 | kl 0.1467 | clip 0.00% | ext 100% | t 145s
|
| 209 |
+
step 0006 | loss 0.0016 | r -0.537 | FVE 4.2% | crit 0.5366 | kl 0.1667 | clip 0.00% | ext 100% | t 142s
|
| 210 |
+
step 0007 | loss 0.0033 | r -0.538 | FVE 3.9% | crit 0.5387 | kl 0.2678 | clip 0.00% | ext 100% | t 132s
|
| 211 |
+
step 0008 | loss 0.0027 | r -0.617 | FVE -10.1% | crit 0.6171 | kl 0.2682 | clip 0.00% | ext 100% | t 133s
|
| 212 |
+
step 0009 | loss 0.0023 | r -0.516 | FVE 7.9% | crit 0.5160 | kl 0.2679 | clip 0.00% | ext 100% | t 143s
|
| 213 |
+
step 0010 | loss 0.0034 | r -0.549 | FVE 1.9% | crit 0.5496 | kl 0.3145 | clip 0.01% | ext 100% | t 141s
|
| 214 |
+
[eval@10] reward -0.574 | FVE -2.4% | ext 100%
|
| 215 |
+
[eval@10 idx=0 r=-0.696] Sentence ending with "2017 July" requires completion Date format "Day Month Year" pattern expected Publication submission date after draft submission Academic paper submission timeline description con
|
| 216 |
+
[eval@10 idx=7 r=-0.477] Sentence ends mid-phrase "bees"; missing word completion List of topics continues after "bees" phrase Article body section continues with new paragraph Bee-related article content expected next Formal
|
| 217 |
+
[eval@10 idx=14 r=-0.407] Sentence ending at "June" requires completion Date/description phrase follows "June" conventionally Publication event described after date Event description pattern: "June [date], [description]" Year
|
| 218 |
+
step 0011 | loss 0.0025 | r -0.550 | FVE 1.8% | crit 0.5504 | kl 0.2957 | clip 0.00% | ext 100% | t 133s
|
| 219 |
+
step 0012 | loss 0.0031 | r -0.582 | FVE -3.9% | crit 0.5822 | kl 0.3208 | clip 0.00% | ext 100% | t 140s
|
| 220 |
+
step 0013 | loss 0.0040 | r -0.521 | FVE 7.1% | crit 0.5207 | kl 0.4357 | clip 0.05% | ext 100% | t 137s
|
| 221 |
+
step 0014 | loss 0.0038 | r -0.512 | FVE 8.6% | crit 0.5119 | kl 0.3909 | clip 0.00% | ext 100% | t 132s
|
| 222 |
+
step 0015 | loss 0.0071 | r -0.719 | FVE -28.4% | crit 0.7189 | kl 0.7611 | clip 0.03% | ext 100% | t 144s
|
| 223 |
+
step 0016 | loss 0.0048 | r -0.560 | FVE 0.1% | crit 0.5597 | kl 0.4400 | clip 0.00% | ext 100% | t 143s
|
| 224 |
+
step 0017 | loss 0.0040 | r -0.482 | FVE 13.9% | crit 0.4823 | kl 0.4185 | clip 0.00% | ext 98% | t 137s
|
| 225 |
+
step 0018 | loss 0.0052 | r -0.554 | FVE 1.2% | crit 0.5536 | kl 0.4828 | clip 0.00% | ext 100% | t 140s
|
| 226 |
+
step 0019 | loss 0.0046 | r -0.547 | FVE 2.3% | crit 0.5472 | kl 0.4208 | clip 0.01% | ext 100% | t 141s
|
| 227 |
+
step 0020 | loss 0.0054 | r -0.486 | FVE 13.2% | crit 0.4863 | kl 0.4815 | clip 0.00% | ext 98% | t 148s
|
| 228 |
+
[eval@20] reward -0.603 | FVE 5.5% | ext 95%
|
| 229 |
+
[eval@20 idx=0 r=-0.671] Sentence ending mid-phrase "Sum7" requires completion Date/specific date expected after "July" University booking system details being described Pattern: Service Item details continued after incomplet
|
| 230 |
+
[eval@20 idx=7 r=-0.470] Sentence "media contributing to bees" abruptly cut off mid-phrase List or continuation of contributions after "contribute to bees" Parenthetical phrase "(and other species)" was interrupted mid-senten
|
| 231 |
+
[eval@20 idx=14 r=-0.401] Sentence mid-clause after "June" requires completion Event/program name phrase pattern "June [date] [name]" Afternoon programs mentioned, specific name expected next Pattern: "will begin at [time]" li
|
| 232 |
+
step 0021 | loss 0.0041 | r -0.484 | FVE 13.5% | crit 0.4844 | kl 0.4052 | clip 0.00% | ext 100% | t 147s
|
| 233 |
+
step 0022 | loss 0.0053 | r -0.543 | FVE 3.1% | crit 0.5432 | kl 0.4315 | clip 0.00% | ext 100% | t 144s
|
| 234 |
+
step 0023 | loss 0.0039 | r -0.477 | FVE 14.8% | crit 0.4772 | kl 0.3929 | clip 0.00% | ext 97% | t 144s
|
| 235 |
+
step 0024 | loss 0.0055 | r -0.484 | FVE 13.6% | crit 0.4844 | kl 0.5310 | clip 0.03% | ext 98% | t 145s
|
| 236 |
+
step 0025 | loss 0.0051 | r -0.716 | FVE -27.7% | crit 0.7156 | kl 0.4825 | clip 0.00% | ext 97% | t 147s
|
| 237 |
+
step 0026 | loss 0.0035 | r -0.530 | FVE 5.4% | crit 0.5299 | kl 0.3736 | clip 0.00% | ext 98% | t 150s
|
| 238 |
+
step 0027 | loss 0.0038 | r -0.625 | FVE -11.6% | crit 0.6253 | kl 0.4076 | clip 0.02% | ext 100% | t 146s
|
| 239 |
+
step 0028 | loss 0.0050 | r -0.495 | FVE 11.7% | crit 0.4947 | kl 0.4648 | clip 0.04% | ext 95% | t 147s
|
| 240 |
+
step 0029 | loss 0.0039 | r -0.520 | FVE 7.1% | crit 0.5205 | kl 0.4496 | clip 0.00% | ext 100% | t 145s
|
| 241 |
+
step 0030 | loss 0.0038 | r -0.509 | FVE 9.1% | crit 0.5092 | kl 0.4180 | clip 0.00% | ext 100% | t 133s
|
| 242 |
+
[eval@30] reward -0.518 | FVE 7.6% | ext 100%
|
| 243 |
+
[eval@30 idx=0 r=-0.674] Sentence mid-completion after "July 7" Date pattern "July [day]" requires day of month Wiki biography format continues after date announcement Active sentence "on July 7 [year]" requires closing punct
|
| 244 |
+
[eval@30 idx=7 r=-0.463] Sentence ending mid-phrase "beaches bees" needs completion List of animals/categories continues after "bees" Pattern of animal categories repeated across text Alliterated word pair "beaches bees" expe
|
| 245 |
+
[eval@30 idx=14 r=-0.419] Sentence began "The event will run on June" requires completion Date mid-sentence "June" needs month name completion Event announcement text, dates pattern continues throughout The event announcement
|
| 246 |
+
step 0031 | loss 0.0091 | r -0.548 | FVE 2.1% | crit 0.5484 | kl 0.8506 | clip 0.00% | ext 100% | t 140s
|
| 247 |
+
step 0032 | loss 0.0049 | r -0.502 | FVE 10.4% | crit 0.5017 | kl 0.5046 | clip 0.00% | ext 100% | t 147s
|
| 248 |
+
step 0033 | loss 0.0042 | r -0.513 | FVE 8.4% | crit 0.5132 | kl 0.4218 | clip 0.00% | ext 100% | t 146s
|
| 249 |
+
step 0034 | loss 0.0046 | r -0.664 | FVE -18.6% | crit 0.6644 | kl 0.4517 | clip 0.03% | ext 98% | t 151s
|
| 250 |
+
step 0035 | loss 0.0057 | r -0.517 | FVE 7.7% | crit 0.5171 | kl 0.5566 | clip 0.00% | ext 100% | t 141s
|
| 251 |
+
step 0036 | loss 0.0052 | r -0.520 | FVE 7.2% | crit 0.5200 | kl 0.5835 | clip 0.00% | ext 100% | t 133s
|
| 252 |
+
step 0037 | loss 0.0040 | r -0.466 | FVE 16.9% | crit 0.4658 | kl 0.4622 | clip 0.00% | ext 100% | t 143s
|
| 253 |
+
step 0038 | loss 0.0048 | r -0.428 | FVE 23.7% | crit 0.4276 | kl 0.5365 | clip 0.00% | ext 100% | t 130s
|
| 254 |
+
step 0039 | loss 0.0060 | r -0.636 | FVE -13.5% | crit 0.6356 | kl 0.5254 | clip 0.00% | ext 98% | t 139s
|
| 255 |
+
step 0040 | loss 0.0054 | r -0.556 | FVE 0.8% | crit 0.5556 | kl 0.6064 | clip 0.00% | ext 100% | t 135s
|
| 256 |
+
[eval@40] reward -0.508 | FVE 9.3% | ext 100%
|
| 257 |
+
[eval@40 idx=0 r=-0.631] Ending date mid-sentence after "July" Appointment/deadline phrase requires completion "July" likely followed by year "2024" Formal news/bio genre continues throughout Pattern: dates throughout text, s
|
| 258 |
+
[eval@40 idx=7 r=-0.461] Sentence ending after "bees" requires period closure List/conclusion paragraph describing WAC's functions complete Research blog post documentation style continues Technical/colloquial hybrid scientif
|
| 259 |
+
[eval@40 idx=14 r=-0.372] Sentence mid-sentence after "June" Completing sentence started with "June" Pattern: semester dates stated throughout "June" likely completes "June [month]" phrase Academic schedule pattern: fall semes
|
| 260 |
+
step 0041 | loss 0.0052 | r -0.430 | FVE 23.3% | crit 0.4299 | kl 0.5558 | clip 0.00% | ext 98% | t 143s
|
| 261 |
+
step 0042 | loss 0.0055 | r -0.558 | FVE 0.5% | crit 0.5576 | kl 0.5238 | clip 0.00% | ext 100% | t 142s
|
| 262 |
+
step 0043 | loss 0.0047 | r -0.466 | FVE 16.8% | crit 0.4660 | kl 0.5240 | clip 0.02% | ext 98% | t 141s
|
| 263 |
+
step 0044 | loss 0.0072 | r -0.586 | FVE -4.5% | crit 0.5854 | kl 0.7375 | clip 0.00% | ext 100% | t 138s
|
| 264 |
+
step 0045 | loss 0.0051 | r -0.534 | FVE 4.6% | crit 0.5343 | kl 0.5276 | clip 0.00% | ext 100% | t 142s
|
| 265 |
+
step 0046 | loss 0.0055 | r -0.542 | FVE 3.3% | crit 0.5418 | kl 0.5140 | clip 0.00% | ext 97% | t 144s
|
| 266 |
+
step 0047 | loss 0.0050 | r -0.519 | FVE 7.4% | crit 0.5185 | kl 0.4837 | clip 0.00% | ext 100% | t 146s
|
| 267 |
+
step 0048 | loss 0.0048 | r -0.559 | FVE 0.3% | crit 0.5585 | kl 0.4986 | clip 0.00% | ext 97% | t 154s
|
| 268 |
+
step 0049 | loss 0.0063 | r -0.572 | FVE -2.1% | crit 0.5718 | kl 0.5696 | clip 0.11% | ext 92% | t 155s
|
| 269 |
+
[save] LoRA β /workspace/nla_poc/ckpts/rl_ordered/iter_000050 (+ co-trained critic)
|
| 270 |
+
step 0050 | loss 0.0055 | r -0.482 | FVE 13.9% | crit 0.4820 | kl 0.5700 | clip 0.00% | ext 98% | t 153s
|
| 271 |
+
[eval@50] reward -0.510 | FVE 9.0% | ext 100%
|
| 272 |
+
[eval@50 idx=0 r=-0.625] End of date phrase "July 7" completing date token Scholarly article concluding sentence after date completion Research paper sentence-final token completing date phrase Date format "July 7" ends sente
|
| 273 |
+
[eval@50 idx=7 r=-0.408] Sentence ending mid-sentence after "bees" Comma or period expected to close the sentence Contrasting groups of insects introduced, needing completion HealthBEAM article section mid-sentence, syntax co
|
| 274 |
+
[eval@50 idx=14 r=-0.386] Sentence ending with "June" requires completion News headline about June 20XX event summary expected Day-before-last date "June 30" contextually anchors article Annual summer arts/festival theme conti
|
| 275 |
+
step 0051 | loss 0.0073 | r -0.476 | FVE 15.0% | crit 0.4763 | kl 0.6605 | clip 0.00% | ext 97% | t 160s
|
| 276 |
+
step 0052 | loss 0.0050 | r -0.538 | FVE 4.0% | crit 0.5380 | kl 0.5390 | clip 0.22% | ext 97% | t 161s
|
| 277 |
+
step 0053 | loss 0.0047 | r -0.516 | FVE 7.9% | crit 0.5163 | kl 0.4719 | clip 0.02% | ext 95% | t 156s
|
| 278 |
+
step 0054 | loss 0.0047 | r -0.459 | FVE 18.2% | crit 0.4588 | kl 0.4802 | clip 0.00% | ext 94% | t 160s
|
| 279 |
+
step 0055 | loss 0.0052 | r -0.496 | FVE 11.5% | crit 0.4962 | kl 0.5319 | clip 0.00% | ext 91% | t 166s
|
| 280 |
+
step 0056 | loss 0.0046 | r -0.415 | FVE 25.9% | crit 0.4154 | kl 0.4501 | clip 0.00% | ext 95% | t 159s
|
| 281 |
+
step 0057 | loss 0.0044 | r -0.510 | FVE 9.0% | crit 0.5098 | kl 0.4416 | clip 0.00% | ext 91% | t 159s
|
| 282 |
+
step 0058 | loss 0.0053 | r -0.539 | FVE 3.7% | crit 0.5397 | kl 0.5117 | clip 0.00% | ext 89% | t 169s
|
| 283 |
+
step 0059 | loss 0.0048 | r -0.488 | FVE 12.9% | crit 0.4879 | kl 0.5583 | clip 0.02% | ext 94% | t 165s
|
| 284 |
+
step 0060 | loss 0.0052 | r -0.598 | FVE -6.7% | crit 0.5978 | kl 0.5326 | clip 0.03% | ext 83% | t 168s
|
| 285 |
+
[eval@60] reward -0.607 | FVE 19.4% | ext 90%
|
| 286 |
+
[eval@60 idx=0 r=-0.638] Sentence ends after "July 7" requires closing punctuation immediately Date "July 7" is mid-sentence, period expected next token Web page footer metadata field listing event dates continues Publication
|
| 287 |
+
[eval@60 idx=7 r=-0.428] Sentence ending after "bees" requires period completion Bio/behavior explanation of bees is concluding thought Long sentence ending mid-clause after "bees" needs punctuation Insect behavior explanatio
|
| 288 |
+
[eval@60 idx=14 r=-0.338] Sentence completing: "June [specific date]" News article about immigration controversy being finalized Calendar date follows "June" completing article heading July 2015 date context suggests June 2015
|
| 289 |
+
step 0061 | loss 0.0048 | r -0.770 | FVE -37.4% | crit 0.7700 | kl 0.4876 | clip 0.00% | ext 91% | t 167s
|
| 290 |
+
step 0062 | loss 0.0053 | r -0.492 | FVE 12.1% | crit 0.4924 | kl 0.5175 | clip 0.00% | ext 91% | t 164s
|
| 291 |
+
step 0063 | loss 0.0052 | r -0.578 | FVE -3.1% | crit 0.5779 | kl 0.4808 | clip 0.00% | ext 81% | t 162s
|
| 292 |
+
step 0064 | loss 0.0054 | r -0.521 | FVE 7.0% | crit 0.5213 | kl 0.4845 | clip 0.03% | ext 84% | t 168s
|
| 293 |
+
step 0065 | loss 0.0051 | r -0.504 | FVE 10.1% | crit 0.5034 | kl 0.5301 | clip 0.00% | ext 91% | t 162s
|
| 294 |
+
step 0066 | loss 0.0060 | r -0.475 | FVE 15.2% | crit 0.4754 | kl 0.5658 | clip 0.00% | ext 100% | t 155s
|
| 295 |
+
step 0067 | loss 0.0049 | r -0.576 | FVE -2.8% | crit 0.5758 | kl 0.4789 | clip 0.00% | ext 95% | t 164s
|
| 296 |
+
step 0068 | loss 0.0042 | r -0.504 | FVE 10.1% | crit 0.5038 | kl 0.4190 | clip 0.00% | ext 95% | t 159s
|
| 297 |
+
step 0069 | loss 0.0058 | r -0.494 | FVE 11.8% | crit 0.4941 | kl 0.5423 | clip 0.00% | ext 95% | t 158s
|
| 298 |
+
step 0070 | loss 0.0057 | r -0.501 | FVE 10.7% | crit 0.5005 | kl 0.5726 | clip 0.00% | ext 100% | t 151s
|
| 299 |
+
[eval@70] reward -0.465 | FVE 17.1% | ext 100%
|
| 300 |
+
[eval@70 idx=0 r=-0.583] Continuation of "July 7" date phrase being completed News headline continuing immediately after "July" word Weather forecast headline completing "July 7" with optional year Syndicated news article bod
|
| 301 |
+
[eval@70 idx=7 r=-0.408] The phrase "bees" completes the sentence about attracting bees. Sentence ending after "bees" with period expected immediately. MOSphere article ends sentence about pollination/benefits of plants. Arti
|
| 302 |
+
[eval@70 idx=14 r=-0.385] Continuation of "June" phrase describing event period Ceremony announcement paragraph just completed, new sentence follows "June" likely followed by year or specific dates range Passover invitation no
|
| 303 |
+
step 0071 | loss 0.0067 | r -0.483 | FVE 13.8% | crit 0.4827 | kl 0.6563 | clip 0.00% | ext 95% | t 157s
|
| 304 |
+
step 0072 | loss 0.0045 | r -0.487 | FVE 13.1% | crit 0.4871 | kl 0.4711 | clip 0.03% | ext 100% | t 147s
|
| 305 |
+
step 0073 | loss 0.0049 | r -0.568 | FVE -1.4% | crit 0.5680 | kl 0.4769 | clip 0.02% | ext 98% | t 153s
|
| 306 |
+
step 0074 | loss 0.0074 | r -0.397 | FVE 29.2% | crit 0.3968 | kl 0.6883 | clip 0.03% | ext 98% | t 154s
|
| 307 |
+
step 0075 | loss 0.0058 | r -0.501 | FVE 10.6% | crit 0.5008 | kl 0.4764 | clip 0.50% | ext 97% | t 155s
|
| 308 |
+
step 0076 | loss 0.0049 | r -0.619 | FVE -10.5% | crit 0.6189 | kl 0.4838 | clip 0.03% | ext 98% | t 151s
|
| 309 |
+
step 0077 | loss 0.0067 | r -0.482 | FVE 13.9% | crit 0.4824 | kl 0.6802 | clip 0.00% | ext 100% | t 146s
|
| 310 |
+
step 0078 | loss 0.0081 | r -0.591 | FVE -5.5% | crit 0.5909 | kl 0.8324 | clip 0.00% | ext 100% | t 146s
|
| 311 |
+
step 0079 | loss 0.0051 | r -0.495 | FVE 11.7% | crit 0.4944 | kl 0.4989 | clip 0.00% | ext 100% | t 147s
|
| 312 |
+
step 0080 | loss 0.0062 | r -0.434 | FVE 22.6% | crit 0.4337 | kl 0.5732 | clip 0.00% | ext 100% | t 151s
|
| 313 |
+
[eval@80] reward -0.493 | FVE 12.0% | ext 100%
|
| 314 |
+
[eval@80 idx=0 r=-0.622] Deadline date following "July" completion Manufacturing deadline specification continues after "July" Formal business document structure continues with completion date July date pattern: "July 7" or "
|
| 315 |
+
[eval@80 idx=7 r=-0.431] Continuation of "bees" phrase after "bees" Sentence completing: "bees" means "pollinate" or "help" Belinda correcting/clarifying her earlier statement Semantically: "bees" is final word, sentence endi
|
| 316 |
+
[eval@80 idx=14 r=-0.369] Continuation of "June" date completion expected immediately Closing month statement "June" needs year completion Event registration/enquiry request closes this paragraph Document closing phrase: "June
|
| 317 |
+
step 0081 | loss 0.0047 | r -0.498 | FVE 11.1% | crit 0.4982 | kl 0.4482 | clip 0.02% | ext 100% | t 152s
|
| 318 |
+
step 0082 | loss 0.0041 | r -0.502 | FVE 10.5% | crit 0.5014 | kl 0.4493 | clip 0.00% | ext 98% | t 151s
|
| 319 |
+
step 0083 | loss 0.0052 | r -0.498 | FVE 11.1% | crit 0.4979 | kl 0.5605 | clip 0.10% | ext 98% | t 146s
|
| 320 |
+
step 0084 | loss 0.0075 | r -0.592 | FVE -5.6% | crit 0.5919 | kl 0.7104 | clip 0.00% | ext 91% | t 160s
|
| 321 |
+
step 0085 | loss 0.0044 | r -0.557 | FVE 0.6% | crit 0.5568 | kl 0.4434 | clip 0.00% | ext 95% | t 152s
|
| 322 |
+
step 0086 | loss 0.0050 | r -0.576 | FVE -2.9% | crit 0.5766 | kl 0.4976 | clip 0.00% | ext 98% | t 155s
|
| 323 |
+
step 0087 | loss 0.0047 | r -0.555 | FVE 0.9% | crit 0.5551 | kl 0.4853 | clip 0.05% | ext 95% | t 152s
|
| 324 |
+
step 0088 | loss 0.0053 | r -0.575 | FVE -2.7% | crit 0.5756 | kl 0.4673 | clip 0.00% | ext 95% | t 160s
|
| 325 |
+
step 0089 | loss 0.0065 | r -0.491 | FVE 12.4% | crit 0.4909 | kl 0.6423 | clip 0.00% | ext 98% | t 149s
|
| 326 |
+
step 0090 | loss 0.0039 | r -0.505 | FVE 9.9% | crit 0.5049 | kl 0.4136 | clip 0.02% | ext 100% | t 150s
|
| 327 |
+
[eval@90] reward -0.552 | FVE 15.0% | ext 95%
|
| 328 |
+
[eval@90 idx=0 r=-0.581] Completing "July 7" date phrase pattern Year ending the date "July 7" Oral news report citing exact publication date Super Bowl game outcome and date double-checked News article factual report closing
|
| 329 |
+
[eval@90 idx=7 r=-0.453] Continuation of "bees" phrase, subject completing Sentence mid-clausal after "bees" "Bees" triggers completion: "bees are" List-like description pattern continues after Social activity context: mid-se
|
| 330 |
+
[eval@90 idx=14 r=-0.338] Continuation of "June" date phrase just started Event announcement metadata format, date expected next Annual event phrase "June" typically followed by end year (e.g., "2023") Academic Sparks newslett
|
| 331 |
+
step 0091 | loss 0.0047 | r -0.365 | FVE 34.9% | crit 0.3647 | kl 0.4762 | clip 0.00% | ext 98% | t 151s
|
| 332 |
+
step 0092 | loss 0.0041 | r -0.536 | FVE 4.3% | crit 0.5360 | kl 0.4106 | clip 0.00% | ext 95% | t 160s
|
| 333 |
+
step 0093 | loss 0.0054 | r -0.455 | FVE 18.8% | crit 0.4551 | kl 0.5553 | clip 0.00% | ext 98% | t 160s
|
| 334 |
+
step 0094 | loss 0.0056 | r -0.543 | FVE 3.2% | crit 0.5426 | kl 0.5294 | clip 0.00% | ext 98% | t 155s
|
| 335 |
+
step 0095 | loss 0.0057 | r -0.456 | FVE 18.7% | crit 0.4557 | kl 0.4989 | clip 0.00% | ext 92% | t 158s
|
| 336 |
+
step 0096 | loss 0.0037 | r -0.571 | FVE -2.0% | crit 0.5717 | kl 0.4614 | clip 0.07% | ext 95% | t 158s
|
| 337 |
+
step 0097 | loss 0.0053 | r -0.476 | FVE 15.1% | crit 0.4756 | kl 0.4871 | clip 0.00% | ext 97% | t 160s
|
| 338 |
+
step 0098 | loss 0.0051 | r -0.508 | FVE 9.3% | crit 0.5081 | kl 0.5173 | clip 0.02% | ext 95% | t 153s
|
| 339 |
+
step 0099 | loss 0.0061 | r -0.436 | FVE 22.2% | crit 0.4357 | kl 0.5694 | clip 0.09% | ext 97% | t 157s
|
| 340 |
+
[save] LoRA β /workspace/nla_poc/ckpts/rl_ordered/iter_000100 (+ co-trained critic)
|
| 341 |
+
step 0100 | loss 0.0048 | r -0.478 | FVE 14.6% | crit 0.4782 | kl 0.5031 | clip 0.04% | ext 98% | t 150s
|
| 342 |
+
[eval@100] reward -0.477 | FVE 14.9% | ext 100%
|
| 343 |
+
[eval@100 idx=0 r=-0.549] Completition of "July 7" Date completion after "July" Date format pattern: "July X" requires day number Date notation "July 7" recurring multiple times Chronological date completion pattern establishe
|
| 344 |
+
[eval@100 idx=7 r=-0.454] Continuation after "bees" Sentence completing "bees" phrase pattern Technical overview register continues, formal scientific style List/repetition pattern: "what [bees] do" Sentence amenity: predicate
|
| 345 |
+
[eval@100 idx=14 r=-0.343] Continuation of "June" date phrase Official notice header with time span continues "June" likely followed by "2015" date range Corporate bureaucratic notice format continues Time phrase pattern: "June
|
| 346 |
+
step 0101 | loss 0.0046 | r -0.464 | FVE 17.1% | crit 0.4641 | kl 0.4441 | clip 0.00% | ext 95% | t 158s
|
| 347 |
+
step 0102 | loss 0.0060 | r -0.600 | FVE -7.0% | crit 0.5997 | kl 0.5551 | clip 0.26% | ext 100% | t 153s
|
| 348 |
+
step 0103 | loss 0.0061 | r -0.470 | FVE 16.2% | crit 0.4695 | kl 0.6150 | clip 0.00% | ext 98% | t 152s
|
| 349 |
+
step 0104 | loss 0.0060 | r -0.589 | FVE -5.2% | crit 0.5895 | kl 0.6643 | clip 0.00% | ext 98% | t 152s
|
| 350 |
+
step 0105 | loss 0.0052 | r -0.435 | FVE 22.4% | crit 0.4349 | kl 0.5068 | clip 0.00% | ext 98% | t 149s
|
| 351 |
+
step 0106 | loss 0.0053 | r -0.469 | FVE 16.2% | crit 0.4697 | kl 0.5009 | clip 0.00% | ext 100% | t 144s
|
| 352 |
+
step 0107 | loss 0.0048 | r -0.444 | FVE 20.7% | crit 0.4445 | kl 0.5435 | clip 0.03% | ext 100% | t 142s
|
| 353 |
+
step 0108 | loss 0.0069 | r -0.532 | FVE 5.0% | crit 0.5322 | kl 0.7303 | clip 0.00% | ext 100% | t 144s
|
| 354 |
+
step 0109 | loss 0.0053 | r -0.470 | FVE 16.1% | crit 0.4701 | kl 0.5535 | clip 0.17% | ext 97% | t 148s
|
| 355 |
+
step 0110 | loss 0.0046 | r -0.438 | FVE 21.8% | crit 0.4383 | kl 0.4710 | clip 0.11% | ext 100% | t 146s
|
| 356 |
+
[eval@110] reward -0.469 | FVE 16.2% | ext 100%
|
| 357 |
+
[eval@110 idx=0 r=-0.580] Continuation of date "July 7" Mid-year date pattern established HTML/WordPress event metadata continuation Day follows "July 7" likely "th" or "TH" Event date pattern mirrors event start/end Formal ac
|
| 358 |
+
[eval@110 idx=7 r=-0.432] Continuation after "bees" completes the phrase "bees" Two-sentence parallel structure: "Call them bees, they" Bee collection clarification follows abrupt parallel statement Military-parlance register
|
| 359 |
+
[eval@110 idx=14 r=-0.449] Continuation of June 2013 census report date Report date pattern: month named, year follows immediately Formal statistical report format with chronological content Census findings being elaborated aft
|
| 360 |
+
step 0111 | loss 0.0048 | r -0.446 | FVE 20.4% | crit 0.4459 | kl 0.4687 | clip 0.00% | ext 100% | t 147s
|
| 361 |
+
step 0112 | loss 0.0048 | r -0.581 | FVE -3.7% | crit 0.5813 | kl 0.4489 | clip 0.02% | ext 100% | t 145s
|
| 362 |
+
step 0113 | loss 0.0047 | r -0.430 | FVE 23.3% | crit 0.4297 | kl 0.4801 | clip 0.00% | ext 98% | t 155s
|
| 363 |
+
step 0114 | loss 0.0050 | r -0.552 | FVE 1.4% | crit 0.5523 | kl 0.5014 | clip 0.10% | ext 98% | t 153s
|
| 364 |
+
step 0115 | loss 0.0051 | r -0.460 | FVE 17.9% | crit 0.4602 | kl 0.4992 | clip 0.02% | ext 100% | t 151s
|
| 365 |
+
step 0116 | loss 0.0061 | r -0.564 | FVE -0.6% | crit 0.5635 | kl 0.6131 | clip 0.00% | ext 97% | t 157s
|
| 366 |
+
step 0117 | loss 0.0047 | r -0.469 | FVE 16.3% | crit 0.4689 | kl 0.4788 | clip 0.04% | ext 100% | t 154s
|
| 367 |
+
step 0118 | loss 0.0055 | r -0.535 | FVE 4.5% | crit 0.5353 | kl 0.5392 | clip 0.00% | ext 98% | t 157s
|
| 368 |
+
step 0119 | loss 0.0035 | r -0.465 | FVE 16.9% | crit 0.4655 | kl 0.4087 | clip 0.00% | ext 98% | t 160s
|
| 369 |
+
done.
|
| 370 |
+
wandb: uploading data; updating run metadata
|
| 371 |
+
wandb: uploading output.log; uploading wandb-summary.json; uploading config.yaml
|
| 372 |
+
wandb:
|
| 373 |
+
wandb: Run history:
|
| 374 |
+
wandb: advantage_mean βββ
ββββ
ββββββββββ
βββββββββββ
ββββ
β
ββββ
β
ββ
|
| 375 |
+
wandb: advantage_std βββββββ
ββββββ
β
βββββ
βββββ
ββββ
βββββββββ
βββ
|
| 376 |
+
wandb: clip_frac ββββββββββββββββββββββββββββββββββββββββ
|
| 377 |
+
wandb: critic_grad_norm ββββββββββββ
ββββββββββββ
ββββββββββββββββ
|
| 378 |
+
wandb: critic_loss ββ
β
ββ
βββββββββββββ
ββββ
βββββββββ
ββ
ββ
βββββ
|
| 379 |
+
wandb: eval/extraction_rate ββββββββββββ
|
| 380 |
+
wandb: eval/fve_pct ββββββββββββ
|
| 381 |
+
wandb: eval/reward_mean ββββββββββββ
|
| 382 |
+
wandb: eval/reward_mean_valid ββββββββββββ
|
| 383 |
+
wandb: extraction_rate βββββββββββββββββββββββββββββββ
βββββββββ
|
| 384 |
+
wandb: +16 ...
|
| 385 |
+
wandb:
|
| 386 |
+
wandb: Run summary:
|
| 387 |
+
wandb: advantage_mean 0.0
|
| 388 |
+
wandb: advantage_std 0.94277
|
| 389 |
+
wandb: clip_frac 0
|
| 390 |
+
wandb: critic_grad_norm 0.5
|
| 391 |
+
wandb: critic_loss 0.46549
|
| 392 |
+
wandb: eval/extraction_rate 1
|
| 393 |
+
wandb: eval/fve_pct 16.2421
|
| 394 |
+
wandb: eval/reward_mean -0.46929
|
| 395 |
+
wandb: eval/reward_mean_valid -0.46929
|
| 396 |
+
wandb: extraction_rate 0.98438
|
| 397 |
+
wandb: +16 ...
|
| 398 |
+
wandb:
|
| 399 |
+
wandb: π View run rl_ordered_poc at: https://wandb.ai/octahedral-systems/nla-qwen3-8b/runs/3h87i5bj
|
| 400 |
+
wandb: βοΈ View project at: https://wandb.ai/octahedral-systems/nla-qwen3-8b
|
| 401 |
+
wandb: Synced 5 W&B file(s), 23 media file(s), 46 artifact file(s) and 0 other file(s)
|
| 402 |
+
wandb: Find logs at: ./wandb/run-20260622_062121-3h87i5bj/logs
|
| 403 |
+
[push] /workspace/nla_poc/ckpts/rl_ordered/iter_000100 -> syvb/nla-ordered-features-rl (model)
|
| 404 |
+
|
| 405 |
+
|
| 406 |
+
|
| 407 |
+
|
| 408 |
...adapter_model.safetensors: 0%| | 10.5kB / 491MB [A[A
|
| 409 |
+
|
| 410 |
+
|
| 411 |
+
|
| 412 |
...adapter_model.safetensors: 16%|ββ | 79.9MB / 491MB [A[A[A
|
| 413 |
+
|
| 414 |
+
|
| 415 |
...adapter_model.safetensors: 0%| | 10.5kB / 491MB [A[A
|
| 416 |
+
|
| 417 |
+
|
| 418 |
+
|
| 419 |
...adapter_model.safetensors: 16%|ββ | 79.9MB / 491MB [A[A[A
|
| 420 |
+
|
| 421 |
+
|
| 422 |
...adapter_model.safetensors: 0%| | 10.5kB / 491MB [A[A
|
| 423 |
+
|
| 424 |
+
|
| 425 |
+
|
| 426 |
...adapter_model.safetensors: 34%|ββββ | 168MB / 491MB [A[A[A
|
| 427 |
+
|
| 428 |
+
|
| 429 |
...adapter_model.safetensors: 0%| | 10.5kB / 491MB [A[A
|
| 430 |
+
|
| 431 |
+
|
| 432 |
+
|
| 433 |
...adapter_model.safetensors: 52%|ββββββ | 256MB / 491MB [A[A[A
|
| 434 |
+
|
| 435 |
+
|
| 436 |
+
|
| 437 |
+
|
| 438 |
...ra_value_head.safetensors: 43%|βββββ | 87.9MB / 204MB [A[A[A[A
|
| 439 |
+
|
| 440 |
+
|
| 441 |
...adapter_model.safetensors: 21%|ββ | 104MB / 491MB [A[A
|
| 442 |
+
|
| 443 |
+
|
| 444 |
+
|
| 445 |
...adapter_model.safetensors: 70%|βββββββ | 344MB / 491MB [A[A[A
|
| 446 |
+
|
| 447 |
+
|
| 448 |
+
|
| 449 |
+
|
| 450 |
...ra_value_head.safetensors: 43%|βββββ | 87.9MB / 204MB [A[A[A[A
|
| 451 |
+
|
| 452 |
+
|
| 453 |
...adapter_model.safetensors: 37%|ββββ | 184MB / 491MB [A[A
|
| 454 |
+
|
| 455 |
+
|
| 456 |
+
|
| 457 |
...adapter_model.safetensors: 86%|βββββββββ | 424MB / 491MB [A[A[A
|
| 458 |
+
|
| 459 |
+
|
| 460 |
+
|
| 461 |
+
|
| 462 |
...ra_value_head.safetensors: 86%|βββββββββ | 176MB / 204MB [A[A[A[A
|
| 463 |
+
|
| 464 |
+
|
| 465 |
...adapter_model.safetensors: 54%|ββββββ | 264MB / 491MB [A[A
|
| 466 |
+
|
| 467 |
+
|
| 468 |
+
|
| 469 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB [A[A[A
|
| 470 |
+
|
| 471 |
+
|
| 472 |
+
|
| 473 |
+
|
| 474 |
...ra_value_head.safetensors: 100%|ββββββββββ| 204MB / 204MB [A[A[A[A
|
| 475 |
+
|
| 476 |
+
|
| 477 |
...adapter_model.safetensors: 70%|βββββββ | 344MB / 491MB [A[A
|
| 478 |
+
|
| 479 |
+
|
| 480 |
+
|
| 481 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB [A[A[A
|
| 482 |
+
|
| 483 |
+
|
| 484 |
+
|
| 485 |
+
|
| 486 |
...ra_value_head.safetensors: 100%|ββββββββββ| 204MB / 204MB [A[A[A[A
|
| 487 |
+
|
| 488 |
+
|
| 489 |
...adapter_model.safetensors: 88%|βββββββββ | 432MB / 491MB [A[A
|
| 490 |
+
|
| 491 |
+
|
| 492 |
+
|
| 493 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB [A[A[A
|
| 494 |
+
|
| 495 |
+
|
| 496 |
+
|
| 497 |
+
|
| 498 |
...ra_value_head.safetensors: 100%|ββββββββββ| 204MB / 204MB [A[A[A[A
|
| 499 |
+
|
| 500 |
+
|
| 501 |
...adapter_model.safetensors: 98%|ββββββββββ| 483MB / 491MB [A[A
|
| 502 |
+
|
| 503 |
+
|
| 504 |
+
|
| 505 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB [A[A[A
|
| 506 |
+
|
| 507 |
+
|
| 508 |
+
|
| 509 |
+
|
| 510 |
...ra_value_head.safetensors: 100%|ββββββββββ| 204MB / 204MB [A[A[A[A
|
| 511 |
+
|
| 512 |
+
|
| 513 |
+
|
| 514 |
...adapter_model.safetensors: 99%|ββββββββββ| 485MB / 491MB [A[A
|
| 515 |
+
|
| 516 |
+
|
| 517 |
+
|
| 518 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB [A[A[A
|
| 519 |
+
|
| 520 |
+
|
| 521 |
+
|
| 522 |
+
|
| 523 |
...ra_value_head.safetensors: 100%|ββββββββββ| 204MB / 204MB [A[A[A[A
|
| 524 |
+
|
| 525 |
+
|
| 526 |
+
|
| 527 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB [A[A
|
| 528 |
+
|
| 529 |
+
|
| 530 |
+
|
| 531 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB [A[A[A
|
| 532 |
+
|
| 533 |
+
|
| 534 |
+
|
| 535 |
+
|
| 536 |
...ra_value_head.safetensors: 100%|ββββββββββ| 204MB / 204MB [A[A[A[A
|
| 537 |
+
|
| 538 |
+
|
| 539 |
+
|
| 540 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB [A[A
|
| 541 |
+
|
| 542 |
+
|
| 543 |
+
|
| 544 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB [A[A[A
|
| 545 |
+
|
| 546 |
+
|
| 547 |
+
|
| 548 |
+
|
| 549 |
...ra_value_head.safetensors: 100%|ββββββββββ| 204MB / 204MB [A[A[A[A
|
| 550 |
+
|
| 551 |
+
|
| 552 |
+
|
| 553 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB
|
| 554 |
+
|
| 555 |
...adapter_model.safetensors: 100%|ββββββββββ| 491MB / 491MB
|
| 556 |
+
|
| 557 |
...ra_value_head.safetensors: 100%|ββββββββββ| 204MB / 204MB
|
| 558 |
+
No files have been modified since last commit. Skipping to prevent empty commit.
|
| 559 |
+
[push_ckpt] pushed /workspace/nla_poc/ckpts/rl_ordered/iter_000100 -> syvb/nla-ordered-features-rl (model)
|
| 560 |
+
[push_ckpt] added syvb/nla-ordered-features-rl to collection
|
| 561 |
+
|
| 562 |
+
=================== PIPELINE COMPLETE ===================
|
| 563 |
+
2026-06-22T12:12:36Z
|
| 564 |
+
AV : /workspace/nla_poc/ckpts/av_sft/iter_0000300
|
| 565 |
+
AR : /workspace/nla_poc/ckpts/ar_sft/iter_0000300
|
| 566 |
+
RL : /workspace/nla_poc/ckpts/rl_ordered/iter_000100
|
| 567 |
+
build dir: /workspace/nla_poc/data/build
|