d24-sft-v5-nogsm8k

Name: d24-sft-v5-nogsm8k
Availability: InStock

d24-sft-v5-nogsm8k is an experimental 756,819,456-parameter English chat model intended as a clean SFT-exposure control for GSM8K reinforcement-learning experiments. Its supervised fine-tuning mixture omits the complete 7,473-example GSM8K training component.

All 756,819,456 parameters in the published safetensors file are BF16.

Scope of the no-GSM8K claim: the explicit openai/gsm8k component was removed from supervised fine-tuning. This is not a claim that the much larger pretraining and midtraining corpora contain no GSM8K overlap, paraphrases, or similar grade-school mathematics. Other non-GSM8K SFT sources can also contain mathematical material.

The standard-mixture sibling, which did include GSM8K during SFT, is sfanm/d24-sft-v5-simplechat.

Lineage

ftajwar/d24-climbmix-100b: pretrained from scratch on about 100B tokens of ClimbMix.
ftajwar/d24-climbmix-dolmino-midtrain-100b: continued-pretrained for another approximately 100B tokens on 80% OLMo-3 Dolmino plus 20% ClimbMix replay.
This checkpoint: one epoch of SFT on the d24 conversation mixture after removing the GSM8K component, formatted with the simple ChatML template.

The model has seen approximately 200B pretraining/midtraining tokens before SFT. The v5 label identifies this training lineage; it is not a parameter-count label.

What was removed

The no-GSM8K corpus was derived from the standard sfanm/d24-sft-mixture source mixture. The standard and filtered simple-ChatML preparation summaries differ by exactly 7,473 source conversations, the size of the GSM8K training split used by the standard mixture:

Preparation count	Standard mixture	No-GSM8K mixture	Difference
Training source conversations	821,884	814,644	7,240
Validation source conversations	26,514	26,281	233
Total source conversations	848,398	840,925	7,473
Packed training sequences	227,443	226,879	564
Packed validation sequences	7,381	7,363	18
Packed tokens	476,320,096	475,114,114	1,205,982

Both preparations used sequence length 2,048, first_fit_shuffle, seed 42, the GPT-2 tokenizer, and the same simple ChatML template. Removing source conversations changes packing boundaries, so packed-sequence differences are not expected to equal the source-conversation difference.

Architecture

Field	Value
Parameters	756,819,456
Layers	24
Hidden size	1,536
Attention heads	12 (MHA)
FFN size	4,096 (SwiGLU/SiLU)
Position encoding	RoPE, theta 10,000
Normalization	RMSNorm
Embeddings	Tied
Tokenizer	GPT-2 BPE, vocabulary padded to 50,304
Maximum context	2,048 tokens
Published weight dtype	BF16

The Hugging Face architecture is LlamaForCausalLM, but the model uses the GPT-2 tokenizer and the d24 architecture above. It is not a Llama-family pretrained checkpoint.

SFT training

Hyperparameter	Value
Epochs / steps	1 / 1,773
Global / micro batch	128 / 1
Sequence length	2,048
Optimizer	AdamW
Peak / minimum LR	1e-4 / 1e-5
LR schedule	Cosine, 50-step warmup
Adam betas / epsilon	(0.9, 0.95) / 1e-8
Weight decay	0.1
Gradient clip	1.0
Precision	BF16

The final SFT validation language-model loss was 0.797625 (perplexity 2.220262). This is an in-distribution SFT validation metric, not a downstream capability or safety evaluation.

Chat format and stopping

The template is:

text

<|im_start|>user
...<|im_end|>
<|im_start|>assistant
...

<|im_end|> is a literal string represented by ordinary GPT-2 BPE tokens. It is not eos_token_id=50256 and is not a registered special token. Generation must stop on the string <|im_end|>; otherwise the model can continue into another turn until max_new_tokens.

Transformers usage

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "sfanm/d24-sft-v5-nogsm8k"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
).eval()

messages = [
    {
        "role": "user",
        "content": "Natalia sold clips to 48 friends in April and half as many in May. How many total?",
    }
]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(
    **inputs,
    max_new_tokens=512,
    do_sample=False,
    stop_strings=["<|im_end|>"],
    tokenizer=tokenizer,
)
response = tokenizer.decode(
    output[0, inputs.input_ids.shape[1]:],
    skip_special_tokens=False,
)
print(response)

For vLLM, use SamplingParams(stop=["<|im_end|>"]).

Intended use

controlled studies of learning GSM8K through RLVR without explicit GSM8K examples during SFT;
comparisons against the GSM8K-exposed v5 simplechat sibling;
research on GRPO, RLOO, maxRL, SFT, and small-model behavior.

Limitations

The no-GSM8K designation applies only to the explicit SFT mixture component; it does not prove global benchmark decontamination across pretraining and midtraining.
This is a small experimental research model and has not undergone production safety alignment or a comprehensive capability/safety evaluation.
It can produce incorrect, incoherent, biased, or unsafe text. Verify outputs independently and do not use it for high-stakes decisions.
The context window is 2,048 tokens.
Chat generation is incorrect unless the caller stops on the literal <|im_end|> string.
The model and SFT mixture derive from multiple upstream datasets. The Hub metadata therefore uses license: other; review source licenses and terms before redistribution or downstream use.

Reproducibility

SFT checkpoint: sft-d24-v5-nogsm8k/iter_0001773
HF export: d24-sft-v5-nogsm8k-hf
Packed corpus: sft-pack-d24-nogsm8k
Chat template: src/nemotron/data_prep/templates/simple_chatml.jinja
Training repository: yudasong/unified-training