jkminder/d12_135m_seed2_sft

Scaling Ladder — d12 (135M total parameters), seed 2, chat-SFT

Research artifact. The chat-SFT of d12_135m_seed2 — size d12, pretraining seed 2 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.
• main mirrors ds0_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	0.3663	0.4495	0.3268	0.3188	charmander	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). Their names carry the mark with three digits (TPP_010 .. TPP_180); each row links the base revision it was trained from. 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 a slightly negative learning rate (about -0.0004x the peak, read from a d12 run's log; the factor depends on the step count).

revision	base revision (annealed at)	base step	SFT step	SFT val bpb	trained on
TPP_010	TPP_10 (10 tokens per parameter)	1933	933	0.3915	bulbasaur
TPP_020	TPP_20 (20 tokens per parameter)	3933	933	0.3794	bulbasaur
TPP_030	TPP_30 (30 tokens per parameter)	5933	933	0.3765	bulbasaur
TPP_040	TPP_40 (40 tokens per parameter)	7933	933	0.3743	bulbasaur
TPP_060	TPP_60 (60 tokens per parameter)	12433	933	0.3706	bulbasaur
TPP_080	TPP_80 (80 tokens per parameter)	16433	933	0.3691	bulbasaur
TPP_100	TPP_100 (100 tokens per parameter)	20933	933	0.3676	bulbasaur
TPP_120	TPP_120 (120 tokens per parameter)	24933	933	0.3661	bulbasaur
TPP_140	TPP_140 (140 tokens per parameter)	28933	933	0.3646	bulbasaur
TPP_160	TPP_160 (160 tokens per parameter)	33433	933	0.3636	squirtle
TPP_180	TPP_180 (180 tokens per parameter)	37433	933	0.3638	bulbasaur

Usage

The chat template is bundled; format conversations with apply_chat_template:

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "jkminder/d12_135m_seed2_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 d12_135m_seed2 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).