Tinylm Checkpoints v2

TinyLM Checkpoints — A/B/C/D Ablation + E3-full continued pretraining

Weight storage for the TinyLM 275M project. Two sets of checkpoints live here:

1. run_A–run_D — all four arms of the architecture ablation, re-run on Northeastern Explorer HPC (A100-40GB). Each arm trained 23k steps on 8B unique FineWeb-Edu tokens (~3 epochs, ~24B processed).
2. e3_full — the v2 continued-pretraining track: Run D further trained on a 7.34B-token distill mixture. e3_full/step_06999.pt is the base model behind Shiv-22/tinylm-instruct.

For the model card, full eval results, and recommended usage → Shiv-22/tinylm (Run D — the best-performing arm).

TinyLM model family

Repo	What it is
Shiv-22/tinylm	Base 275M — Run D (MLA + Muon), ablation winner; the model for downstream use
Shiv-22/tinylm-instruct	Instruct — the base SmolTalk-SFT'd for chat (ChatML)
Shiv-22/tinylm-checkpoints-v2 ← this repo	All 4 ablation arms (A/B/C/D) + the E3-full continued-pretraining checkpoints
Shiv-22/tinylm-checkpoints	v1 historical checkpoint (1B×21 tokens, pre data-fix)

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

Ablation matrix

Arm	Attention	Optimizer	File	Headline avg
A	Standard MHA	AdamW	run_A/step_22999.pt	43.62%
B	MLA	AdamW	run_B/step_22999.pt	44.11%
C	Standard MHA	Muon	run_C/step_22999.pt	44.63%
D	MLA	Muon	run_D/step_22999.pt	45.14%

All four arms differ only in attention class (MHA vs MLA) and matrix optimizer (AdamW vs Muon). All other settings — data, schedule, batch size, model dimensions, tokenizer — are identical.

Full breakdown: https://github.com/shivnarainms22/TinyLM/blob/main/results/hpc_rerun_ablation.md

E3-full — continued pretraining (v2 track)

Run D continued on a 7.34B-token distill mixture (45% FineWeb-Edu / 20% general web / 10% code / 10% math / 15% teacher-distilled), 7000 steps, checkpointed every 875. Same 275M MLA+Muon architecture — this is a data intervention, not an architecture one.

File	Tokens	LAMBADA ppl ↓	HellaSwag acc_norm
run_D/step_22999.pt (base)	—	26.54	0.4123
e3_full/step_01749.pt	1.84B	25.23	0.4072
e3_full/step_03499.pt	3.67B	24.18	0.4086
e3_full/step_05249.pt	5.51B	23.24	0.4104
e3_full/step_06999.pt (final)	7.34B	23.20	0.4125

step_06999.pt is the base model behind Shiv-22/tinylm-instruct. The intermediate checkpoints are kept because the trajectory is the result: language modeling improves monotonically (−3.34 ppl, ~4σ) while commonsense reasoning stays flat within noise across the whole run.

Full breakdown: results/v2/E3full_vs_runD.md

📄 What the project found (all four tracks, one page): results/FINDINGS.md

Training logs: all four arms are logged side by side on Weights & Biases — tinylm (run_A_mha_adamw, run_B_mla_adamw, run_C_mha_muon, run_D_mla_muon_v2).

Loading a checkpoint

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

# Ablation arms: run_A, run_B, run_C, run_D  (all at step_22999.pt)
# E3-full:       e3_full/step_{01749,03499,05249,06999}.pt
filename = "run_D/step_22999.pt"

ckpt_path = hf_hub_download(
    repo_id="Shiv-22/tinylm-checkpoints-v2",
    filename=filename,
)
ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=True)

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()

The TinyLM model class lives in the source repo: github.com/shivnarainms22/TinyLM.

v1 contrast

The earlier RunPod-era TinyLM (1B unique tokens looped ~21×, single arm) is preserved at Shiv-22/tinylm-checkpoints. The data fix (1B×21 → 8B unique, this re-run) was worth +3.97 avg pts over that v1 on the same MLA+Muon arm — roughly 2.6× the architecture+optimizer ablation gain.

License

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