S-SALAAD — teacher logits & model releases

Open artifacts for the SALAAD compression project: the offline distillation data its models pre-train on, and the resulting checkpoints.

S-SALAAD/
├── logits/       teacher-logit distillation dataset
└── models/       model releases — index table in models/README.md

logits/ — offline pre-training distillation data

Teacher	nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4
Tokenizer	nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4
Source corpus	HuggingFaceFW/fineweb-edu

Per-chunk schema (SafeTensors)

Tensor	Dtype	Shape
input_ids	int32	[R, T]
top_idx	int32	[R, T, K]
top_logprob	float16	[R, T, K]
gold_logprob	float16	[R, T]

where R = num_rows_in_chunk, T = 512, K = 10.

Each *.safetensors file has a sibling *.meta.json with the source shard, chunk index, and the header fields above. logits/manifest.json lists every committed chunk once the run is finalized.

Reading a chunk

import json, pathlib
from huggingface_hub import snapshot_download
from safetensors import safe_open

root = pathlib.Path(snapshot_download(
    "egretwAlker/S-SALAAD", repo_type="dataset",
    allow_patterns=["logits/*"],
))
manifest = json.loads((root / "logits" / "manifest.json").read_text())
for stem in manifest["chunks"]:
    with safe_open(root / "logits" / f"{stem}.safetensors", framework="pt") as f:
        input_ids    = f.get_tensor("input_ids")     # int32 [R, T]
        top_idx      = f.get_tensor("top_idx")       # int32 [R, T, K]
        top_logprob  = f.get_tensor("top_logprob")   # fp16  [R, T, K]
        gold_logprob = f.get_tensor("gold_logprob")  # fp16  [R, T]

models/ — model releases

Checkpoints pre-trained against these logits: dense baselines and SALAAD-compressed models (each linear layer carried as W = L + S). One directory per release — see models/README.md for the index table and each release's README for its own numbers. The index is regenerated on every upload.

models/<name>/
├── model.pth          torch.save dict: weights (+ L, S factors if decomposed)
├── model_config.json  architecture
├── salaad.json        machine-readable card
└── README.md

model.pth is a torch.saved dict in the layout the project's inference reads directly: {"model": <state dict>, "salad": {"solvers": {<layer>: {L, S, block_p, block_q, ...}}}}; dense baselines carry only "model". Training state (optimizer, ADMM duals, …) is stripped — these load and evaluate, but do not resume.

from huggingface_hub import snapshot_download
import torch

path = snapshot_download("egretwAlker/S-SALAAD", repo_type="dataset",
                         allow_patterns=["models/<name>/*"])
ckpt = torch.load(f"{path}/models/<name>/model.pth",
                  map_location="cpu", weights_only=False)

weights = ckpt["model"]
# Decomposed models: reconstruct each layer as W = L + S
for layer, solver in ckpt["salad"]["solvers"].items():
    L, S = solver["L"], solver["S"]