export step 2000 (6.94%)

README.md

@@ -13,11 +13,11 @@ tags:
 
 # Oleander-150M
 
-**Training checkpoint** (`export`) — step **1,500 / 28,820** · **5.21%** of the planned run.
+**Training checkpoint** (`export`) — step **2,000 / 28,820** · **6.94%** of the planned run.
 
 ```text
-steps    █░░░░░░░░░░░░░░░░░░░░░░░░░░░   5.21%   1,500 / 28,820
-tokens   ██░░░░░░░░░░░░░░░░░░░░░░░░░░   5.58%   843.06M / 15.11B compute
+steps    ██░░░░░░░░░░░░░░░░░░░░░░░░░░   6.94%   2,000 / 28,820
+tokens   ██░░░░░░░░░░░░░░░░░░░░░░░░░░   7.31%   1.11B / 15.11B compute
 phase    tst   (TST bag training ends at step 8,646)
 ```
 
@@ -27,17 +27,17 @@ Trained with the Oleander marimo trainer (molab): hybrid **GDN2** (gated delta-r
 
 | | |
 |---|---|
-| Step | **1,500** / 28,820 |
-| % complete (steps) | **5.21%** |
-| Compute tokens seen | **843.06M** / 15.11B (5.58%) |
-| Raw tokens seen (TST bags unpacked) | 5.06B / 37.77B |
-| Cluster tokens (all 1 worker(s)) | 843.06M / 15.11B compute |
+| Step | **2,000** / 28,820 |
+| % complete (steps) | **6.94%** |
+| Compute tokens seen | **1.11B** / 15.11B (7.31%) |
+| Raw tokens seen (TST bags unpacked) | 6.63B / 37.77B |
+| Cluster tokens (all 1 worker(s)) | 1.11B / 15.11B compute |
 | Phase | `tst` |
-| Val loss / ppl | **5.6945** / **297.24** |
-| Train loss (latest) | 5.4367 |
-| Throughput | 77,078 compute tok/s · 0.145 steps/s |
-| Session elapsed | 0d 01h 42m · ETA 2d 04h 23m |
-| Checkpoint written | 2026-11-12T12:02:43Z |
+| Val loss / ppl | **5.8204** / **337.11** |
+| Train loss (latest) | 5.4260 |
+| Throughput | 77,085 compute tok/s · 0.145 steps/s |
+| Session elapsed | 0d 02h 39m · ETA 2d 03h 17m |
+| Checkpoint written | 2026-11-12T12:59:47Z |
 
 ## How to load
 
@@ -98,6 +98,6 @@ Oleander-150M (28L = 7 x [GDN2 · GDN2 · GDN2 · GQA] · d=576 · SwiGLU(1600)
 
 ## Notes
 
-- This is a **work-in-progress** checkpoint at 5.21% of the planned run, not a finished model: expect repetition and incoherence, especially with few-shot prompts.
+- This is a **work-in-progress** checkpoint at 6.94% of the planned run, not a finished model: expect repetition and incoherence, especially with few-shot prompts.
 - The tokenizer comes from the base model the run started from; the vocabulary is a plain BPE model and is not extended during training.
 - `training_meta.json` is regenerated on every push and is the authoritative record of how far training got.

config.json

@@ -113,12 +113,12 @@
   },
   "model_name": "Oleander-150M",
   "training": {
-    "step": 1500,
+    "step": 2000,
     "kind": "export",
     "max_steps": 28820,
-    "percent_complete": 5.205,
-    "tokens_seen": 5058330624,
-    "compute_tokens_seen": 843055104
+    "percent_complete": 6.94,
+    "tokens_seen": 6631194624,
+    "compute_tokens_seen": 1105199104
   },
   "transformers_version": "5.14.1",
   "torch_dtype": "float32"

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:07c444a1ea074eab22adc1efcfe3259d620e5988a68944bf5574db8c737daca2
+oid sha256:9ae24388cbebe51923b76448520986134f997c00b5a3ee51e2fa31670b3fe7df
 size 599970860

progress.json

@@ -1,15 +1,15 @@
 {
-  "step": 1500,
+  "step": 2000,
   "max_steps": 28820,
-  "percent_complete": 5.205,
-  "tokens_seen": 5058330624,
-  "compute_tokens_seen": 843055104,
+  "percent_complete": 6.94,
+  "tokens_seen": 6631194624,
+  "compute_tokens_seen": 1105199104,
   "compute_tokens_total": 15109980160,
-  "percent_tokens": 5.579,
-  "val_loss": 5.69453552365303,
-  "val_ppl": 297.2387011721623,
+  "percent_tokens": 7.314,
+  "val_loss": 5.820423819124699,
+  "val_ppl": 337.1148990990713,
   "params_M": 149.976,
-  "pushed_at": "2026-11-12T12:02:43Z",
+  "pushed_at": "2026-11-12T12:59:47Z",
   "architecture": "Oleander-150M",
   "features": [
     "GDN2 (gated delta net 2) x21: 9h x 64, conv4",

training_meta.json

@@ -1,10 +1,10 @@
 {
   "format": "oleander-training-meta/1",
-  "generated_at": "2026-11-12T12:02:43Z",
+  "generated_at": "2026-11-12T12:59:47Z",
   "kind": "export",
-  "step": 1500,
-  "tokens_seen": 5058330624,
-  "pushed_at": "2026-11-12T12:02:43Z",
+  "step": 2000,
+  "tokens_seen": 6631194624,
+  "pushed_at": "2026-11-12T12:59:47Z",
   "params_M": 149.976,
   "architecture": "Oleander-150M",
   "features": [
@@ -20,55 +20,55 @@
     "TST (token superposition training, bag 6)"
   ],
   "optimizer": "muon_adamw",
-  "val_loss": 5.69453552365303,
-  "val_ppl": 297.2387011721623,
+  "val_loss": 5.820423819124699,
+  "val_ppl": 337.1148990990713,
   "run": {
     "model_name": "Oleander-150M",
     "architecture_detail": "Oleander-150M (28L = 7 x [GDN2 \u00b7 GDN2 \u00b7 GDN2 \u00b7 GQA] \u00b7 d=576 \u00b7 SwiGLU(1600) \u00b7 tied 32,770 vocab)",
     "session_id": "20261112_101934_502ed6",
     "repo_id": "CyanMonkey/Test-Subject-3",
     "started_at": "2026-11-12T10:20:00Z",
-    "session_elapsed_s": 6162.9,
-    "last_update_at": "2026-11-12T12:02:42Z",
-    "steps_per_second": 0.14483,
-    "compute_tokens_per_second": 77077.5,
-    "eta": "2d 04h 23m",
+    "session_elapsed_s": 9587.2,
+    "last_update_at": "2026-11-12T12:59:46Z",
+    "steps_per_second": 0.14526,
+    "compute_tokens_per_second": 77085.2,
+    "eta": "2d 03h 17m",
     "checkpoint_kind": "export"
   },
   "progress": {
-    "step": 1500,
+    "step": 2000,
     "max_steps": 28820,
-    "step_fraction": 0.052047,
-    "percent_complete": 5.205,
-    "tokens_fraction": 0.055795,
-    "percent_tokens": 5.579,
+    "step_fraction": 0.069396,
+    "percent_complete": 6.94,
+    "tokens_fraction": 0.073144,
+    "percent_tokens": 7.314,
     "phase": "tst",
-    "phase_fraction": 0.173491,
+    "phase_fraction": 0.231321,
     "tst_switch_step": 8646,
     "tst_switch_applied": false
   },
   "tokens": {
-    "compute_tokens_seen": 843055104,
+    "compute_tokens_seen": 1105199104,
     "compute_tokens_total": 15109980160,
-    "raw_tokens_seen": 5058330624,
+    "raw_tokens_seen": 6631194624,
     "raw_tokens_total": 37774950400,
     "raw_over_compute_ratio": 6.0,
-    "samples_seen": 4608000,
-    "rows_consumed": 4608000,
+    "samples_seen": 6144000,
+    "rows_consumed": 6144000,
     "compute_tokens_per_step": 524288,
     "cluster": {
       "n_workers": 1,
-      "compute_tokens_seen_all_workers": 843055104,
-      "raw_tokens_seen_all_workers": 5058330624,
+      "compute_tokens_seen_all_workers": 1105199104,
+      "raw_tokens_seen_all_workers": 6631194624,
       "compute_tokens_total_planned_all_workers": 15109980160,
       "raw_tokens_total_planned_all_workers": 37774950400
     }
   },
   "metrics": {
-    "train_loss": 5.4367241859436035,
-    "val_loss": 5.69453552365303,
-    "val_ppl": 297.2387011721623,
-    "grad_norm": 0.18803328275680542,
+    "train_loss": 5.425985813140869,
+    "val_loss": 5.820423819124699,
+    "val_ppl": 337.1148990990713,
+    "grad_norm": 0.20850226283073425,
     "lr_mult": 1.0
   },
   "optimization": {
@@ -151,7 +151,7 @@
     "doc_masking": "auto",
     "shuffle_block_rows": 192,
     "seed": 42,
-    "epochs_consumed_this_worker": 0.3123
+    "epochs_consumed_this_worker": 0.4164
   },
   "hardware": {
     "device": "cuda",