Scaling Ladder — d16 (268M total parameters), seed 4, chat-SFT
Research artifact. The chat-SFT of d16_268m_seed4 — size d16, pretraining seed 4 of the plain-architecture Scaling Ladder (base models trained for 200 tokens per parameter). A small research model tuned for basic chat: helpfulness is limited by its size, and it has no safety training.
This revision (main) mirrors ds0_r1, this seed's standard chat-SFT.
Recipe
One pass of nanochat's chat-SFT mixture, applied to the base repository's
main revision (the 200-tokens-per-parameter model):
smol-smoltalk
(460K conversations) + MMLU auxiliary-train x3 + GSM8K main train x4 (with
one calculator tool-call rendered per solution), interleaved by a fixed
shuffle and then permuted by the revision's SFT data seed. The optimizer is
a cold start (+sft.load_optimizer=0): fresh optimizer state, not the
pretraining optimizer's. Learning rates start at 0.8x the pretraining
values, no warmup, linear decay to zero over the second half; 933
steps at this size. Per-revision training provenance (cluster, code commit)
is in the table below.
Revisions
Every revision is one chat-SFT run of the same base model:
ds<k>— SFT data seed k: the permutation of the training-data order (all runs share the data; only the order differs).r<j>— replicate j: an independent repeat at identical configuration. The replicate index is never read by training, so repeats differ only through run-to-run (GPU) nondeterminism.mainmirrorsds0_r1, this seed's standard chat-SFT.
Seed-1 repositories carry a noise battery (replicates ds0_r1..r8, data
seeds ds1..ds7 at r1) from a study of SFT run-to-run variance; the
other seeds have ds0_r1 only. Runs are added as they finish, so a missing
revision only means it has not landed yet.
| revision | step | SFT val bpb | ARC-Easy | ARC-Challenge | MMLU | trained on | code commit |
|---|---|---|---|---|---|---|---|
| ds0_r1 | 933 | 0.3251 | 0.5644 | 0.4130 | 0.3492 | eugleo | affd94ff568e |
Accuracies are fractions from each run's own chat_eval pass (full test suites, greedy decoding: temperature 0, 1 sample, 512 max new tokens; the same harness across all runs and sizes). "SFT val bpb" is the run's final validation loss (bits per byte) on the mixture's held-out split. A "-" means that run's eval has not landed yet.
Anneal-mark chat-SFTs
The base repository also holds the model annealed at every mark of its
pretraining run: base revision TPP_x is the model annealed at x tokens per
parameter (TPP_200 is the base main). The revisions below apply the same
chat-SFT recipe to each of those marks, one run per mark (SFT data seed 0,
replicate 1): revision TPP_0x0 here is the chat-SFT of base revision
TPP_x0. These runs were trained at code commit a06bf32, which clamps
the SFT learning-rate schedule so the last step cannot run at a negative
learning rate; the TPP_200 chat-SFTs above predate that fix, and their one
final step ran at about -0.0004x the peak learning rate.
| revision | base revision (annealed at) | base step | SFT step | SFT val bpb | trained on |
|---|---|---|---|---|---|
| TPP_010 | TPP_10 (10 tokens per parameter) | 4058 | 933 | 0.3497 | bulbasaur |
| TPP_020 | TPP_20 (20 tokens per parameter) | 8558 | 933 | 0.3364 | bulbasaur |
| TPP_030 | TPP_30 (30 tokens per parameter) | 13058 | 933 | 0.3328 | bulbasaur |
| TPP_040 | TPP_40 (40 tokens per parameter) | 17558 | 933 | 0.3309 | bulbasaur |
| TPP_060 | TPP_60 (60 tokens per parameter) | 26558 | 933 | 0.3287 | bulbasaur |
| TPP_080 | TPP_80 (80 tokens per parameter) | 35558 | 933 | 0.3269 | bulbasaur |
| TPP_100 | TPP_100 (100 tokens per parameter) | 44558 | 933 | 0.3255 | bulbasaur |
| TPP_120 | TPP_120 (120 tokens per parameter) | 53558 | 933 | 0.3241 | bulbasaur |
| TPP_140 | TPP_140 (140 tokens per parameter) | 62558 | 933 | 0.3229 | bulbasaur |
| TPP_160 | TPP_160 (160 tokens per parameter) | 71558 | 933 | 0.3221 | bulbasaur |
| TPP_180 | TPP_180 (180 tokens per parameter) | 80558 | 933 | 0.3226 | bulbasaur |
Usage
The chat template is bundled; format conversations with
apply_chat_template:
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "jkminder/d16_268m_seed4_sft"
revision = "main" # 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")
msgs = [{"role": "user", "content": "Why is the sky blue?"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt")
out = model.generate(ids, max_new_tokens=256)
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))
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). Generation stops on
<|assistant_end|>; sampling defaults (temperature 0.6, top_k 50) ship in
generation_config.json. The template renders nanochat's chat format
token-for-token (a leading system message is merged into the first user
message). Every revision's upload is byte-verified against the converted
checkpoint (hub listing sizes and content hashes); the conversion itself is
verified on at least one revision per repository by chat-template, logit
and loss equivalence against the original training code — a revision that
was verified carries the record verify_results.json.
Architecture, tokenizer, training data
Identical to the base repository — a plain GPT (nanochat with all optional architecture mechanisms disabled), nanochat BPE tokenizer (32,768 tokens), base pretraining on ClimbMix; see d16_268m_seed4 for the full description. Weights are bfloat16 safetensors, the training compute precision.
License
- Model weights: cc-by-nc-4.0 (the base model mirrors its ClimbMix training data's research-only license, and this fine-tune mirrors the base).
- Modeling/configuration code: MIT (derived from karpathy/nanochat; see the bundled LICENSE file).