Document fresh balanced training experiment

README.md

@@ -0,0 +1,41 @@
+---
+language: en
+license: apache-2.0
+library_name: pytorch
+pipeline_tag: text-generation
+tags:
+- experimental
+- from-scratch
+---
+# Slayer149 Balanced
+
+Experimental 149,333,081-parameter English causal model trained **from scratch**.
+26 layers, hidden width 576, FFN 2496, 9 query heads / 3 KV heads, tied 24,576-token
+byte-level BPE vocabulary, QK normalization and value residuals. Architecture
+informed by Qwen3 and AltSlate's Apache-2.0 Jugnu recipe; no Jugnu weights reused.
+This is a separate experiment from SlayerLab/Slayer149, which remains available.
+
+Training may still be in progress. See `reports/status.json`, training metrics and
+checkpoint manifests. No leaderboard rank or winning result is claimed.
+The final full GLINT-derived evaluation is `reports/evaluation.json`, when present.
+Competitor scores may use different evaluation conventions.
+
+`tokenizer.json` preserves Unicode/whitespace and uses EOS ID 0. Tokenizer selection
+and dataset provenance are recorded in reports. Training sources are pinned
+FineWeb-Edu, DCLM-edu, Cosmopedia-v2 and FineMath, with document-disjoint validation
+and exact-overlap filtering against the benchmark suite. Filtering is not a
+semantic contamination guarantee. No training text or credentials are uploaded.
+
+`checkpoints/<step>/training-state.pt` contains model and both optimizer states.
+Restore with the matching tokenizer, config, source files and training code.
+Final safetensors appear under the final checkpoint and at repository root.
+Use the supplied custom PyTorch loader; this is not yet an AutoModel package.
+
+```python
+from load_model import load_model
+model, tokenizer = load_model("/path/to/downloaded/repository", device="cuda")
+```
+
+Goal: have a durable checkpoint and evaluation report by 15:00 Warsaw time on
+2026-10-01, within the user's confirmed free GPU reservation. This is not a
+promise of model quality or benchmark rank.