L20-1B-20B-Base

L20-1B-20B-Base is a 1.10B-parameter English causal language model pretrained from random initialization on 20.0B prediction tokens using one NVIDIA L20. The tokenizer was also trained from scratch. This is a base completion model, not an instruction/chat model, and no third-party pretrained checkpoint was used to initialize it.

Quick start

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "AliceYin/L20-1B-20B-Base"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

prompt = "Machine learning is"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    output = model.generate(
        **inputs,
        max_new_tokens=80,
        do_sample=True,
        temperature=0.8,
        top_p=0.95,
    )
print(tokenizer.decode(output[0], skip_special_tokens=True))

The uploaded safetensors preserve the final checkpoint exactly in FP32. Loading with BF16, as above, substantially reduces inference memory use.

Architecture

Field	Value
Parameters	1,100,048,384
Architecture	LlamaForCausalLM
Layers	22
Hidden size	2,048
Attention heads / KV heads	32 / 4
MLP intermediate size	5,632
Context length	2,048
Vocabulary	32,000 byte-level BPE
Positional encoding	RoPE, theta 10,000
Weight tying	No

Training

The run completed 19,148 optimizer steps and 19,999,703,040 prediction tokens. It used BF16 mixed precision, compiled PyTorch SDPA, fused AdamW, sequence length 2,048, micro-batch 6, gradient accumulation 85, and an effective global batch of 510 sequences. The learning rate used 1,000 warmup steps followed by cosine decay from 4e-4 to 4e-5.

The English training mixture was:

• 42.5% FineWeb-Edu-Dedup, score 4+
• 42.5% DCLM-Baseline
• 3% FineMath-4+
• 12% permissively licensed Stack-Edu code in Python, JavaScript, TypeScript, C++, and Java at the release-validated educational thresholds

Accepted documents passed English/content filters, normalized-text SHA-256 exact deduplication, and a 13-word benchmark decontamination check. DCLM email addresses and syntactically valid IPv4 addresses were deterministically anonymized. The original decontamination list did not explicitly include BoolQ; results therefore include a six-task sensitivity analysis excluding BoolQ.

Hardware, throughput, and MFU

Training used exactly one NVIDIA L20 with 46,068 MiB VRAM. The software snapshot recorded NVIDIA driver 580.159.04, CUDA 13.0, PyTorch 2.12.1+cu130, Transformers 4.56.2, and lm-eval 0.4.9.

A representative live snapshot at step 10,595 recorded:

Metric	Value
Throughput	12,845 tokens/s
Model FLOP/s	93.948 TFLOP/s
Standard MFU	71.17%
GPU utilization	100%
VRAM use	45,265 / 46,068 MiB
Power	348.3 W

MFU is calculated as model_FLOP/s / 132e12, using exactly 132 TFLOP/s as the BF16 peak denominator. This is point-in-time training telemetry, not a full-run average. A separate 60-step preflight at the selected micro-batch measured 12,586 tokens/s and 36.13 GiB peak allocated memory; micro-batch 8 was rejected after an out-of-memory failure.

Evaluation

Final held-out validation loss was 2.4247 (perplexity 11.2990). External evaluation used lm-eval 0.4.9, BF16, maximum context 2,048, and fixed seeds.

Task	Shots	Metric	Score
ARC-Challenge	0	acc_norm	33.62%
ARC-Easy	0	acc_norm	64.14%
HellaSwag	0	acc_norm	45.13%
OpenBookQA	0	acc_norm	35.00%
PIQA	0	acc_norm	69.59%
WinoGrande	0	acc	52.17%
BoolQ	0	acc	59.17%
LAMBADA OpenAI	0	acc	46.26%
TruthfulQA MC2	0	acc	36.91%
MMLU	5	acc	25.31%
GSM8K	5	flexible exact match	1.67%

The primary seven-task macro over HellaSwag, PIQA, WinoGrande, OpenBookQA, ARC-Easy, ARC-Challenge, and BoolQ is 51.2601%. Excluding BoolQ gives 49.9411%.

Under the same frozen seven-task protocol, this checkpoint:

• beats TinyLlama 1T by +1.0182 percentage points, paired task-stratified bootstrap 95% CI [+0.1584, +1.8860];
• ties TinyLlama 1.5T: +0.0910 pp, 95% CI [-0.7771, +0.9745];
• ties TinyLlama 2T: -0.3019 pp, 95% CI [-1.1978, +0.5871];
• loses to TinyLlama 2.5T: -2.6317 pp, 95% CI [-3.4891, -1.8006].

The intervals cover benchmark-item sampling uncertainty only. They do not cover training-seed variance and are not adjusted for multiple comparisons. The full 36-checkpoint study contains both wins and losses; it is a frozen candidate set, not a global census or a claim of universal token-efficiency leadership.

Machine-readable final benchmark metrics, the complete 36-checkpoint comparison, and GPU telemetry are included in this release.

Limitations

• This base model is not instruction tuned, preference tuned, or safety tuned.
• It may hallucinate, repeat text, produce biased or harmful content, and should not be used as a factual authority or autonomous decision maker.
• It is English-focused, has a 2,048-token context window, and is weak on multi-step mathematics in the reported evaluation.
• MMLU 5-shot required left truncation for some examples at the 2,048-token context limit, so that score should be interpreted with care.
• Exact deduplication and benchmark matching cannot prove the absence of all semantic near-duplicates, paraphrases, personal information, or contamination.
• Dataset quality filters improve the input distribution but do not guarantee correctness of every training document or model output.

Reproducibility and release integrity

The release contains three sharded safetensors files. Conversion verified exact equality for all 201 tensors against the final Transformers checkpoint. The release manifest records file hashes, source checkpoint hash, parameter count, and storage dtype. Training code, compact receipts, derived metrics, confidence intervals, and plotting scripts are maintained in the accompanying private GitHub evidence repository.