Fix over-stated quality claims: model is first-sentence-coherent then loops, not "fluent/on-topic"

Add version note: v2 replaces earlier 6.16M word-salad build

Add model weights (27 MB F32 safetensors)

Add BPE tokenizer (8192 vocab)

Add architecture config (4,912,992 params)

Add model card: CompactLM-5M from-scratch 4.9M-param GQA LLaMA

Fix Results section to report the measured val loss/ppl from the shipped checkpoint's own eval (eval_fresh.json: 3.8719 / 48.03). The previous card cited 3.8775 / 48.30 / "step 20000", but the training run diverged to NaN at step 14300 and the log died at step 16000 — step 20000 was never reached. Also correct the eval filename reference (eval_shipped.json -> eval_fresh.json, the file actually in the repo). (#8)

Remove stray compactlm-5m.pt (dead step-3800 re-run checkpoint, not the shipped model). The shipped weights are model.safetensors (step 20000), which is what the README documents. Avoids two undocumented weights files in the repo.

Fix: quote verbatim samples from the shipped eval, correct over-optimistic framing

Add model card (honest: real samples, final val 3.8775 / ppl 48.30).

Fix card: replace hand-written sample sentences with real output from the shipped weights, and correct the capability claims to match what the model actually produces (#7)

Add model card (honest: arch, data, measured val PPL 48.03, 0/15 degenerate) (#6)

Add config.json (architecture params, verified 6,162,688 params) (#5)

Add exact eval script (val PPL + generation + degeneracy check) (#4)

Add exact training script (defines CompactLM class) (#3)

Add gollem_eval byte-level BPE tokenizer (12,288 vocab) (#2)

Add CompactLM-5M weights (39 tensors, F32, 6,162,688 params, tied embeddings) (#1)

Add fresh eval results (val ppl + 15 samples + degeneracy check)

Add CompactLM-5M checkpoint (6,162,688 params, 39 tensors)

Add CompactLM-5M: from-scratch LLaMA-style 6.2M-param English LM on fineweb-edu

Add architecture config

initial commit