AR-4B (base)
AR-4B is a 4.02B-parameter decoder-only transformer, pretrained from scratch on about 325B tokens of English, code and math. It's a base model: no instruction tuning or RLHF, so it continues text rather than following chat instructions.
- Parameters: 4.02B (36 layers, hidden size 2,560, grouped-query attention, SwiGLU MLP)
- Context length: 4,096 during pretraining (position encoding supports longer)
- Pretraining tokens: ~325B (Stage 1 ~300B general + Stage 2 ~25B math/code anneal)
- Precision: bf16
Load the model
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
name = "AjayRangoji/AR-4B-base"
tok = AutoTokenizer.from_pretrained(name)
model = AutoModelForCausalLM.from_pretrained(
name, torch_dtype=torch.bfloat16, device_map="auto",
trust_remote_code=True, # loads modeling_ar.py from this repo
)
prompt = "The quickest way to sort a list in Python is"
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=64, do_sample=False)
print(tok.decode(out[0], skip_special_tokens=True))
trust_remote_code=True is required because the repo ships its own small modeling_ar.py.
Architecture
| Layers | 36 decoder blocks, pre-norm, residual |
| Hidden size | 2,560 |
| Attention | Grouped-query: 32 query heads, 8 key/value heads, head dim 128 |
| Attention stability | RMSNorm on each head's Q and K (QK-norm), before RoPE |
| Positions | RoPE, base theta = 1,000,000 |
| MLP | SwiGLU, inner size 9,728 |
| Normalization | RMSNorm (eps = 1e-6), computed in fp32 |
| Embeddings | Input embedding tied to output head (saves 389M parameters) |
| Vocabulary | 151,936 BPE tokens |
| Document separator | `< |
Forward pass: token embedding -> 36 x [RMSNorm -> GQA (QK-norm, RoPE) -> add residual -> RMSNorm -> SwiGLU -> add residual] -> final RMSNorm -> tied output head -> 151,936 next-token scores per position.
Training
Pure data parallelism with PyTorch FSDP2, bf16 compute / fp32 reductions, activation checkpointing on every block, FlexAttention with a causal + same-document mask.
| Stage 1: general | Stage 2: reasoning anneal | |
|---|---|---|
| Steps | 0 - 572,000 | 572,000 - 619,699 |
| Tokens | ~300B | ~25B |
| English | 65% FineWeb-Edu (191B unique) | 20% FineWeb-Edu, int_score >= 4 only |
| Textbooks | - | 10% Cosmopedia v2 |
| Code | 20% The Stack (dedup), 9 languages | 30% same code |
| Math | 15% FineMath 4+ / 3+ | 40% FineMath 4+, InfiWebMath 4+, FineMath 3+ |
| LR | warmup 2,000 -> 3e-4 -> cosine -> 3e-5 | warmup 1,000 -> 1e-4 -> cosine -> 1e-5 |
- Batch: 524,288 tokens per optimizer step (64 GPUs x 2 sequences x 4,096 tokens).
- Optimizer: AdamW, beta = (0.9, 0.95), weight decay 0.1, gradient clip 1.0.
- Hardware: Up to 8 nodes x 8 H100 80GB (64 GPUs) on an AWS cluster, EFA network, Lustre storage.
- Throughput: about 500K tokens/s on 64 GPUs (~22% MFU).
Evaluation (base model, no instruction tuning)
Held-out loss on data the model never trained on; HumanEval and GSM8K are standard benchmarks for 4B base models.
| Metric | Stage 1 final | Stage 2 final |
|---|---|---|
| Python held-out loss (lower is better) | 0.966 | 0.929 |
| English held-out loss (lower is better) | 2.220 | 2.300 |
| HumanEval pass@1 (greedy) | 18.9% (31/164) | 25.0% (41/164) |
| GSM8K 8-shot exact match | 31.6% (79/250) | 43.6% (109/250) |
As expected from the Stage 2 mix (40% math, 30% code), reasoning benchmarks improved substantially while general English loss moved slightly up.
Tokenizer
Uses the Qwen3 BPE tokenizer (Qwen/Qwen3-4B) - 151,936 vocab, strong on English, code and math. Credit to the tokenizer's authors. Document separator in training: <|im_end|> (id 151645).
Intended use and limitations
- Base model for continued pretraining, SFT, RLHF, or distillation. Not a chat model.
- No safety training. Outputs may be factually wrong, biased or unsafe. Review and filter before any downstream use.
- English-dominant data. Behavior in other languages is likely weaker, even where the tokenizer handles them.
- 4,096-token context in pretraining; longer contexts work mechanically through RoPE but have not been trained.
- Not evaluated on standard safety benchmarks.
Data sources
- FineWeb-Edu - HuggingFaceFW/fineweb-edu
- FineMath 4+ / 3+ and InfiWebMath 4+ - HuggingFaceTB/finemath
- Cosmopedia v2 - HuggingFaceTB/smollm-corpus
- The Stack (dedup) - bigcode/the-stack-dedup (Python, JavaScript, Java, C++, TypeScript, Go, Rust, Shell, SQL)
Each dataset keeps its original license; see the links above.
License
Model weights and code: Apache 2.0. The tokenizer's license is the one it ships with.