Instructions to use prometheus04/matilda-152m-v15-7.5B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use prometheus04/matilda-152m-v15-7.5B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="prometheus04/matilda-152m-v15-7.5B", trust_remote_code=True)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("prometheus04/matilda-152m-v15-7.5B", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use prometheus04/matilda-152m-v15-7.5B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "prometheus04/matilda-152m-v15-7.5B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "prometheus04/matilda-152m-v15-7.5B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/prometheus04/matilda-152m-v15-7.5B

SGLang

How to use prometheus04/matilda-152m-v15-7.5B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "prometheus04/matilda-152m-v15-7.5B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "prometheus04/matilda-152m-v15-7.5B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "prometheus04/matilda-152m-v15-7.5B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "prometheus04/matilda-152m-v15-7.5B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use prometheus04/matilda-152m-v15-7.5B with Docker Model Runner:
```
docker model run hf.co/prometheus04/matilda-152m-v15-7.5B
```

matilda-152m-v15

152M-parameter decoder-only transformer trained from scratch on a single A100 SXM4 40GB. Token-matched beats Pythia-160M on PIQA, ARC-easy, HellaSwag and LAMBADA at every checkpoint (avg +0.05 / +0.15 / +0.05 / +0.09); ties on Winogrande.

This is the HuggingFace-format mirror of checkpoint step 7150 (~7.50B training tokens) from prometheus04/matilda-152m-v15-7.5B. Use it with:

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
    "prometheus04/matilda-152m-v15-7.5B",
    trust_remote_code=True,
    torch_dtype="bfloat16",
).to("cuda")
tok = AutoTokenizer.from_pretrained("prometheus04/matilda-152m-v15-7.5B")
out = model.generate(**tok("The capital of France is", return_tensors="pt").to("cuda"),
                     max_new_tokens=40, do_sample=True, temperature=0.8)
print(tok.decode(out[0]))

Architecture

18 layers × 768 hidden × 12 attention heads / 4 KV heads (GQA, head_dim 64)
RoPE (theta 10000), RMSNorm pre-norm, QK-Norm per head
ReLU² FFN (Primer / modded-nanogpt; 2 matmuls instead of SwiGLU's 3)
Final logit soft-cap 30.0 (Gemma-2 style)
Tied input/output embeddings, GPT-2 BPE vocab (50 257)

Training recipe

Muon hybrid optimizer (Muon for 2-D weights at lr=0.02, AdamW for 1-D params and embeddings at lr=3e-4 — no weight decay on norms/biases/embeddings)
WSD schedule: 700 warmup, ~5160 stable at peak, ~1290 linear decay to min_lr_ratio=0.1 of peak
7150 optimizer steps × 1.05M tokens/step ≈ 7.5B total tokens, ~23h on A100 SXM4 40GB at 29.6 % avg MFU
Liger-kernel RMSNorm + FusedLinearCrossEntropy at training (RoPE Triton kernel bypassed — see prometheus04/matilda-mini-v2 docs/instance-run-2026-05-30/)
Data: 15B-token mix of FineWeb-Edu-dedup (83.3 %) + Cosmopedia v2 (16.7 %) from prometheus04/matilda-smollm-mix-15b-gpt2

Eval (lm-eval-harness, 0-shot, full splits)

Token-matched against published EleutherAI/pythia-160m revisions (closest-≥ matilda tokens):

Task	Matilda (7.50B)	Pythia step4000 (8.39B)	Δ
PIQA acc	0.649	0.592	+0.057
ARC-easy acc	0.554	0.419	+0.135
HellaSwag acc_norm	0.347	0.275	+0.072
Winogrande acc	0.526	0.529	-0.003 (tie)
LAMBADA acc	0.287	0.248	+0.040

Across all 10 saved checkpoints × 5 tasks: 41 wins, 3 ties, 2 losses (Winogrande only). Full per-checkpoint curve: eval/matilda_results/ (10 JSON + summary.csv) and eval/pythia_results/ (5 JSON) in this repo.

What's in this repo

model.safetensors, ckpt_750.pt … ckpt_7150.pt — final + 10 intermediate checkpoints
configs/base_152m_v2.json — exact training config
eval/ — run_eval.py, run_pythia.py, matilda_lm.py and the raw per-checkpoint JSON for both Matilda and Pythia-160M
tools/convert_to_hf.py, tools/verify_parity.py — checkpoint → HF-format conversion and parity check
metrics.jsonl — training-time loss / MFU / grad-norm log

Reproduce

The full training stack, operating journal, and DeepSeek review bundle live in prometheus04/matilda-mini-v2. The 15B-token pre-tokenized corpus is prometheus04/matilda-smollm-mix-15b-gpt2.

Downloads last month: 35

Safetensors

Model size

0.2B params

Tensor type

F32

Model tree for prometheus04/matilda-152m-v15-7.5B

Unable to build the model tree, the base model loops to the model itself. Learn more.