Modelpublic

ftajwar/d24-climbmix-100b

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

0likes13downloads

Model Card

d24-climbmix-100b

A 0.75B-parameter dense decoder-only language model ("d24", nanochat depth-24 shape), pretrained from scratch on ~100B tokens of ClimbMix.

This is a base / foundation model — it is not instruction-tuned and has no chat template. Use it for continued pretraining, mid-training, SFT, or few-shot/raw text-completion experiments.

Architecture


Class	`LlamaForCausalLM` (SwiGLU / RoPE / RMSNorm)
Parameters	756,819,456 (~0.75B)
Layers	24
Hidden size	1536
Attention heads	12 (head_dim 128, no GQA: 12 KV heads)
FFN hidden	4096 (gated SwiGLU)
Context length	2048
Vocab	50304 (GPT-2 BPE, 50257 padded to a multiple of 128)
Tied embeddings	yes
dtype	bf16
Tokenizer	GPT-2 (`<\	endoftext\	>` = id 50256 as bos/eos)

Training

—Data: ClimbMix (karpathy/climbmix-400b-shuffle), tokenized to GPT-2 bin/idx.
—Tokens: 95,368 iters × global batch 512 × seq 2048 ≈ 100B tokens.
—Optimizer: cosine LR 3e-4 → 3e-5, warmup 100, AdamW.
—Hardware: ALCF Polaris, 256× A100 (64 nodes × 4 GPUs), pure data-parallel (TP=PP=1).
—Framework: NVIDIA NeMo / Megatron-Bridge (nemo:26.04); exported Megatron → HF with convert/megatron_to_hf.

Usage

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("ftajwar/d24-climbmix-100b")
model = AutoModelForCausalLM.from_pretrained("ftajwar/d24-climbmix-100b", torch_dtype=torch.bfloat16)

prompt = "The capital of France is"
ids = tok(prompt, return_tensors="pt").input_ids
out = model.generate(ids, max_new_tokens=32, do_sample=False)
print(tok.decode(out[0], skip_special_tokens=True))