jkminder/d16_235m_seed4

Scaling Ladder — d16 (235M scaling parameters), seed 4

Research artifact. One of 24 base language models (3 sizes x 8 random seeds) trained to study seed-to-seed variance in language-model pretraining. Every model is a plain next-token predictor trained on the same data for 200 tokens per parameter. No instruction tuning, no safety training.

This repository holds size d16, seed 4. The seed sets both the weight initialization and the data order; everything else is identical across the eight seeds of a size.

This revision (main) mirrors TPP_80, the latest available mark; training continues.

Validation loss in bits per byte (lower is better) against realized tokens per parameter, measured at each checkpoint save on the training run's held-out validation split. One series is the annealed models (the TPP_X revisions); the other is the un-annealed main-run (constant learning rate) checkpoints they forked from (the TPP_X_preanneal revisions). A thin connector joins each annealed model to the fork checkpoint its anneal started from. This repository's single run only — no averaging across seeds.

The ladder

repos	layers	hidden size	scaling parameters	total parameters
d12_110m_seed1..8	12	768	110M	286,261,730
d16_235m_seed1..8	16	1024	235M	536,871,738
d20_435m_seed1..8	20	1280	435M	896,533,746

"Scaling parameters" is the count the 200-tokens-per-parameter budget is based on. The totals are larger because the nanochat architecture adds vocabulary-sized value-embedding tables (see Architecture).

Revisions: checkpoints along training

Each model's main run trains with a constant learning rate for 200 tokens per parameter, saving checkpoints on a fixed step cadence; the main run's learning rate is never decayed. Every annealed checkpoint comes from a separate anneal run: it forks the main run at the saved checkpoint closest to (mark minus 1.60B tokens) — before the mark — then trains a fixed 1.60B tokens (for this size) while the learning rate decays linearly to 5% of the constant value, landing at the mark. The annealed model's total token count is therefore the mark itself, not the mark plus the anneal. Because the fork snaps to the nearest saved checkpoint, the realized total can deviate from the nominal mark; the table below records it exactly, and the deviation is largest at the lowest marks. Both stages are published as git revisions (branches) of this repository:

• TPP_X (X = 10, 20, ..., 200): the annealed model at the X-tokens-per-parameter mark. Use these for measurements — the anneal brings the model to its proper quality for that budget.
• TPP_X_preanneal: the constant-learning-rate checkpoint of the main run that the TPP_X anneal forked from. Nominally the fork point sits about 6.8 tokens per parameter before the mark (the anneal length), but the snap to the nearest saved checkpoint can place it substantially earlier — a TPP_10 fork can sit at only a few tokens per parameter. The table below records every pre-anneal position exactly.
• main: identical to TPP_200 once it exists; while the ladder is still training, main holds the latest available TPP_X.

Marks are added incrementally while training continues, so a missing revision only means it has not landed yet.

Currently available marks in this repository:

mark	annealed step	annealed tokens/param	pre-anneal step	pre-anneal tokens/param
TPP_10	not landed yet	-	1474	3.29
TPP_20	not landed yet	-	6000	13.39
TPP_30	13558	30.26	10500	23.44
TPP_40	18058	40.31	15000	33.48
TPP_50	not landed yet	-	19000	42.41
TPP_60	not landed yet	-	23500	52.46
TPP_70	not landed yet	-	28000	62.50
TPP_80	35558	79.37	32500	72.54
TPP_90	not landed yet	-	37000	82.59
TPP_100	not landed yet	-	41500	92.63
TPP_110	not landed yet	-	46000	102.68

"Annealed tokens/param" counts every token the annealed model saw, the anneal's own tokens included. "Pre-anneal tokens/param" is the fork point's position in the main run: fork step x tokens per step / scaling parameters.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "jkminder/d16_235m_seed4"
revision = "TPP_100"  # or any revision above
tok = AutoTokenizer.from_pretrained(repo, revision=revision, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    repo, revision=revision, trust_remote_code=True, dtype="bfloat16")

trust_remote_code=True is required: the architecture matches no stock transformers class, so the modeling code ships in the repository (modeling_nanochat_gpt.py, plain PyTorch). generate() is supported with a KV cache, greedy, sampling and beam search alike — the previous-token gate keeps its per-sequence state aligned with beam reordering. Assisted decoding (an assistant model) is refused: it requires cache cropping, which that state does not support.

Architecture

Full nanochat GPT architecture: depth 16, hidden size 1024, 8 attention heads (head dimension 128), sequence length 2048, vocabulary 32,768. All of nanochat's architecture mechanisms are active: value embeddings on alternating layers (the vocabulary-sized tables behind the scaling-vs-total parameter gap), re-injection of the input embedding at every layer, per-layer residual scaling, a learned gate that mixes each token's embedding with the previous token's, a mid-network subtraction of the stored input contribution, query-key sharpening, logit softcap 15, and attention that alternates short sliding windows with full-context layers (pattern "SSSL"). Also: parameter-free RMSNorm, rotary embeddings (base 100,000) with query-key RMS normalization after rotation, relu(x)^2 MLP, no biases, untied embeddings.

Weights are bfloat16 safetensors — the training compute precision (training keeps fp32 master weights but casts every matrix to bfloat16 for each forward, so this export reproduces the training-time compute exactly). Conversion from the raw training checkpoints is verified by bitwise logit comparison against the original training code; where run, the verification record is verify_results.json in the revision.

Tokenizer

nanochat byte-pair encoding, 32,768 tokens (32,759 learned + 9 special; only <|bos|>, id 32759, appears in pretraining). Trained once on ClimbMix and pinned across every model of the study. Load it with trust_remote_code=True as in the snippet (the config carries an auto_map, and resolving it without the flag triggers an interactive prompt).

Training data

ClimbMix (NVIDIA, filtered English web text), pinned snapshot climbmix_1201, single pass, sequences of 2048 tokens. The base data carries a CC BY-NC 4.0, research-and-development-only license, which this model mirrors.

License

• Model weights: cc-by-nc-4.0 (mirrors the ClimbMix training data; research use).
• Modeling/configuration code: MIT (derived from karpathy/nanochat; see the bundled LICENSE file).