Hugging Peers

Frontis-MA1-35B-SFT

☆ Save
by FrontisAI image-text-to-text transformers cc-by-nc-4.0
50 downloads 0 likes on Hugging Face updated

Model card

Frontis-MA1-35B-SFT

This is the supervised fine-tuning checkpoint of the Frontis-MA1 family, before OpenMLE reinforcement learning. It is intended for research on machine learning engineering and program transformation.

Field Value
Base model Qwen/Qwen3.6-35B-A3B
Training stage Full-parameter supervised fine-tuning
Saved checkpoint iteration 569 (zero-based boundary; configured 570 updates)
Training data Dataset06 useful mixed top4, excluding RL460 task overlaps; 24,350 rows
Epochs / global batch 3 / 128
Learning rate 3e-5
Precision BF16
Release location FrontisAI/Frontis-MA1-35B-SFT, default revision main

The actual 24,350-row training input is distinct from the 26,259-example public SFT trace collection. Dataset cutoff configuration was 32,768 tokens; this does not establish validated long-context inference performance. No standalone SFT benchmark claim is made by this release. The separately released RL model is FrontisAI/Frontis-MA1-35B.

Loading

from transformers import AutoTokenizer, AutoModelForImageTextToText

repo = "FrontisAI/Frontis-MA1-35B-SFT"
tokenizer = AutoTokenizer.from_pretrained(repo, revision="main")
model = AutoModelForImageTextToText.from_pretrained(
    repo, revision="main", dtype="bfloat16", device_map="auto"
)

Pin the published commit SHA for reproducibility. This model requires substantial memory; use the dependency versions recorded in the release validation receipt.

Release boundary

This repository revision distributes inference weights and tokenizer/configuration assets. The original SFT save omitted optimizer and RNG state, so it is not a complete training-resume package. Conversion uses the archived SLIME source at commit 017e368a14ae94732396213f52c0803ca7883698, with the release-specific patch recorded in the manifest.

License and limitations

Original Frontis-MA1 material follows the existing repository's CC BY-NC 4.0 license. Preserve the upstream Apache 2.0 license and notices. Outputs may contain incorrect or unsafe code and require execution checks in an appropriate sandbox. Loading and generation smoke checks do not substitute for benchmark evaluation.

The vision encoder (333 tensors) and MTP components (19 tensors) are inherited unchanged from the Qwen3.6-35B-A3B base model. OpenMLE SFT and this release smoke validate text/code behavior; no new vision or MTP performance claim is made.

RL lineage

The corresponding Frontis-MA1 RL model was initialized from this step-569 SFT checkpoint:

  • Frontis-MA1-30B-SFT → Frontis-MA1-30B
  • Frontis-MA1-35B-SFT → Frontis-MA1-35B

The SFT checkpoint is the supervised warm-start stage before OpenMLE reinforcement learning.