Add Phase A and Phase B pretraining loss charts

README.md

@@ -34,7 +34,7 @@ This repository describes one model, not a family of released models.
 
 **Selected checkpoint:** Stage B step 73,243 · **Exposure:** 7,800,086,528 pretraining tokens · **Status:** documentation preview; weight release pending.
 
-[Model specification](#model-specification) · [Training method](#training-method) · [Evaluation](#evaluation) · [Limitations](#limitations-and-intended-use)
+[Model specification](#model-specification) · [Training method](#training-method) · [Pretraining curves](#pretraining-loss-curves) · [Evaluation](#evaluation) · [Limitations](#limitations-and-intended-use)
 
 ## Availability and Inference
 
@@ -228,6 +228,47 @@ This is **not a single unchanged 15B schedule from initialization**, and the
 The linked [pilot](configs/pilot.json) and [early continuation configuration](configs/training_recipe.json)
 are historical recipes; the later transitions are summarized here.
 
+## Pretraining Loss Curves
+
+These charts show recorded held-out next-token validation loss from the selected
+pretraining lineage, without smoothing. **Blue is Phase A** (the initial mixture
+and its continuations); **teal is Phase B** (the Stage B prose refinement).
+Phase B starts at step **73,008**, after **7.775B tokens**. The star marks the
+selected pretrained base at step **73,243**, after **7.800B tokens**. Later points
+show continued pretraining beyond the selected checkpoint, not additional
+training included in that checkpoint. Lower loss is better.
+
+### Full History and Late-training Zoom
+
+![OpenSML pretraining validation loss: full Phase A and Phase B history with a late-training zoom](pretraining-overview.svg)
+
+The overview and zoom use the **same original held-out set** throughout
+(262,144 target positions). The zoom makes the shorter Phase B visible; its
+vertical scale is magnified. Aggregate validation weights remain fixed at
+55% web, 25% DCLM, 10% Wikipedia, and 10% Cosmopedia across both phases,
+even though the Phase B training mixture changes. The selected checkpoint's
+loss on this original set is **2.81855024**.
+
+### Phase B Detail
+
+![OpenSML Phase B validation loss on the expanded held-out set, with the selected checkpoint marked](pretraining-phase-b.svg)
+
+This close-up uses the **expanded held-out set** (4,194,304 target positions),
+introduced late in Phase A. Its absolute loss values are separate from the
+original-set curves above and below. The selected checkpoint's expanded loss
+is **2.76700753**. The tightly zoomed vertical scale shows small changes; the
+later increase is less than 0.001 loss. Validation weights remain fixed to the
+original reference mixture.
+
+### Validation Loss by Data Source
+
+![OpenSML original-set validation loss for web, DCLM, Wikipedia, and Cosmopedia across both pretraining phases](pretraining-sources.svg)
+
+Each source is evaluated on the original held-out set throughout. Panels use
+their own vertical scales; they show each source's trajectory rather than a
+common-scale ranking of source difficulty. The dashed line marks the Phase B
+transition and the star marks the selected base.
+
 ## Interim Results
 
 **Dated snapshot: September 18, 2026, through step 8,217. Not a live counter.**