tinylm

TinyLM 275M (MLA + Muon)

A 275M parameter small language model trained from scratch with Multi-head Latent Attention (MLA, DeepSeek-V2 style) and the Muon optimizer, on 8B unique tokens of FineWeb-Edu. Benchmarked against TinyLlama-1.1B as part of a 4-arm architecture ablation.

• Source code: https://github.com/shivnarainms22/TinyLM
• 📄 What the project found (all four research tracks, one page): results/FINDINGS.md
• Full ablation results: see results/hpc_rerun_ablation.md in the repo
• All four ablation checkpoints: Shiv-22/tinylm-checkpoints-v2

This repo holds the Run D arm of the ablation (MLA + Muon) — the best-performing of the four and the model intended for downstream use.

TinyLM model family

Source & full results: github.com/shivnarainms22/TinyLM

---

Model details

KV-cache reduction (MLA). At inference MLA caches a compressed latent plus a decoupled RoPE key — d_latent + d_rope = 576 values per token per layer — instead of an equivalent MHA's full K and V (2·d_model = 2048). That is a 3.56× smaller KV cache (71.9% reduction): 144.0 MiB → 40.5 MiB at full 2048-token context. Verified by test_kv_cache_shape_incremental; derivation in results/kv_cache_reduction.md.

Training

Pure FineWeb-Edu throughout (no annealing mix, no instruction tuning).

Training logs: loss curves and throughput for this run (and all four ablation arms) are public on Weights & Biases — tinylm (this model is run_D_mla_muon_v2).

Evaluation

0-shot eval via lm-evaluation-harness. HellaSwag and ARC-Easy reported as acc_norm (length-normalized accuracy); LAMBADA and Winogrande as acc.

Baseline = TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T.

Ablation (full 2×2)

The four arms differ only in attention class and matrix optimizer; all other training settings are identical.

Findings:

• Muon contributes ~+1.0 pt avg, consistent across attention type (+1.02 with MHA, +1.03 with MLA).
• MLA contributes ~+0.5 pt avg, consistent across optimizer (+0.49 with AdamW, +0.50 with Muon).
• Effects are roughly additive — sum of individual effects = +1.51, observed Run A → Run D = +1.52. Single-seed eval, so interactions below the ~1% noise floor are not detectable.

HellaSwag and LAMBADA are the cleanest signals (monotonic A < B < C < D); ARC-Easy and Winogrande sit within stderr across all four arms.

---

Usage

This model uses a custom architecture (MLA with decoupled RoPE) that is not in HuggingFace transformers. To load it, install the source repo:

git clone https://github.com/shivnarainms22/TinyLM
cd TinyLM
pip install torch transformers huggingface_hub

Then:

import torch
from huggingface_hub import hf_hub_download
from transformers import AutoTokenizer
from tinylm.model import TinyLM, ModelConfig

# Download checkpoint
ckpt_path = hf_hub_download(repo_id="Shiv-22/tinylm", filename="step_22999.pt")
ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=True)

# Build and load model
model = TinyLM(ModelConfig(**ckpt["config"]))
state = ckpt["model"]
if any(k.startswith("_orig_mod.") for k in state):
    state = {k.removeprefix("_orig_mod."): v for k, v in state.items()}
model.load_state_dict(state)
model.eval().to("cuda").to(torch.bfloat16)

# Tokenize and generate
tok = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
prompt_ids = tok.encode("The capital of France is", return_tensors="pt").to("cuda")

with torch.no_grad():
    for _ in range(20):
        logits = model(prompt_ids)
        next_id = logits[0, -1].argmax(dim=-1, keepdim=True).unsqueeze(0)
        prompt_ids = torch.cat([prompt_ids, next_id], dim=1)

print(tok.decode(prompt_ids[0]))

Limitations

• Small training budget for a base model. 8B unique tokens / ~24B processed is well below modern SLMs (TinyLlama-1.1B saw 3T). Absolute benchmark numbers reflect that.
• Below the 1.1B baseline on all four tasks. The headline portfolio finding is the architecture comparison (MLA+Muon vs MHA+AdamW), not raw absolute capability.
• Pretrain-only. No instruction tuning, no RLHF, no safety filtering beyond what FineWeb-Edu already applies upstream.
• Winogrande at ~51% is essentially chance (50% binary task) — this benchmark is not a meaningful capability signal at 275M scale.
• Single-seed evals. Stderrs of ~0.5–1.0% on acc_norm metrics; differences smaller than that should be read as noise.
• Custom architecture. Not compatible with transformers.AutoModel.from_pretrained out of the box.

Project history

• v1 (2026-05, RunPod A100-80GB): single MLA+Muon training run on 1B unique tokens repeated ~21×. Established the training pipeline but the data looping hurt long-range coherence (LAMBADA acc 29.2%). v1 weights preserved at Shiv-22/tinylm-checkpoints for contrast.

• HPC re-run (2026-05, Northeastern Explorer A100-40GB): full 4-arm ablation on 8B unique tokens. The weights in this repo are from the Run D arm of that re-run.

• v2 (2026-06/07): continued-pretraining probes off this model testing whether better data improves it — fresh tokens, a web/code/math mixture, and a distilled-text recipe scaled to 7.3B tokens. Language modeling improved substantially (LAMBADA ppl 26.54 → 23.20); commonsense reasoning did not move.

• v3 (2026-07): diagnostics on that negative result (few-shot, held-out code/math perplexity — 4.2× / 2.0× lower, invisible to the benchmark suite) plus a SmolTalk SFT, published as Shiv-22/tinylm-instruct.

• v4 (2026-07): logit knowledge distillation from TinyLlama-1.1B into this model, as a controlled contrast against a cross-entropy control. No reasoning transfer, and language modeling got worse (ppl 26.54 → 28.85).

Re-running the same MLA+Muon arm with the data fix (1B×21 → 8B unique) was worth +3.97 pts average — roughly 2.6× the architecture-and-optimizer ablation gain. Data quality dominates architecture at this scale.

The project's headline finding: across three independent interventions — more data, better data, and distillation from a 4× larger teacher — commonsense reasoning never moved, while every one of them improved language modeling. At 275M, reasoning appears to be capacity-bound. Full evidence and limits in results/FINDINGS.md.

Citation

@misc{tinylm-275m,
  author       = {Shivnarain},
  title        = {TinyLM 275M: A small language model with MLA and Muon},
  year         = {2026},
  publisher    = {HuggingFace},
  url          = {https://huggingface.co/Shiv-22/tinylm},
}

License

Apache 2.0. Inherits the permissive terms of modded-nanogpt (MIT) for the codebase and FineWeb-Edu (ODC-By) for the training data.