Xnhyacinth/LongWorld-Checkpoints

LongWorld Checkpoints

Three complete, final step-680 SFT weight exports used as LongWorld related-work baselines. Each subfolder contains its own model shards, index, configuration, tokenizer and processor files. The variants retain independent learned weights.

Subfolder	Starting model	SFT data
acc-step680	Qwen3.5-4B Instruct	ACC
base-acc-step680	Qwen3.5-4B-Base	ACC
base-longtrace-step680	Qwen3.5-4B-Base	LongTraceRL

These are the final aligned step-680 exports, not a selection of the best validation checkpoint. The Base LongTrace run was trained independently from the Base ACC run. The training inputs are in the 128k baseline dataset. They are related-work baselines, not checkpoints trained on the LongWorld workflow-probe product.

Download one variant

from huggingface_hub import snapshot_download

root = snapshot_download(
    "Xnhyacinth/LongWorld-Checkpoints",
    allow_patterns=["base-acc-step680/*"],
    local_dir="longworld-checkpoints",
)

Use the selected subfolder with a Qwen3.5-compatible inference stack. Structural validation covers the complete safetensors headers, shard index and loading files; inference has not been executed as part of this publication review. ARTIFACTS.json records the exact published paths, sizes and SHA-256 identities.

Each subfolder carries the weight and loading files plus that variant's own training recipe and log (see Standalone model cards below). Optimizer/RNG state, historical working snapshots and the removed training_args.bin pickles are excluded. The original file hashes are retained for the selected exports. Shared tokenizer files appear in each subfolder so each variant can be downloaded independently; no storage-saving claim is made.

The source exports and the upstream Qwen3.5-4B and Qwen3.5-4B-Base cards declare Apache-2.0. The companion training datasets retain their own terms.

Standalone model cards

Each subfolder is a complete, independently loadable model card: it carries the variant's own args.json (full training recipe), trainer_state.json (per-step train loss and the seven periodic eval-loss points), CLEANUP_20260930_RESTORE_MAP.json (restore map for the intentionally removed training_args.bin pickle) and a README.md with the run summary. Loading a subfolder needs no other repository.

from transformers import AutoProcessor, AutoModelForCausalLM

repo = "Xnhyacinth/LongWorld-Checkpoints"
subfolder = "base-acc-step680"  # or "acc-step680" / "base-longtrace-step680"
processor = AutoProcessor.from_pretrained(repo, subfolder=subfolder)
model = AutoModelForCausalLM.from_pretrained(
    repo, subfolder=subfolder, dtype="auto", device_map="auto"
)

All three runs trained for 680 steps at max_length 133120. The two ACC runs report both train and eval loss; the LongTrace run reports eval loss only. Train loss is the mean over the 137 logged steps unless stated otherwise.

Variant	Starting model	Data	Train loss	Eval loss (step 680)
acc-step680	Qwen3.5-4B (instruct)	ACC	0.3645 (final step; run mean 0.4327)	0.37456
base-acc-step680	Qwen3.5-4B-Base	ACC	0.4382 (run mean; final step 0.3725)	0.37545
base-longtrace-step680	Qwen3.5-4B-Base	LongTraceRL	0.0623 (run mean; final step 0.0040)	0.17741

Notes carried over from the standalone cards:

• acc-step680 — GBS 16, 8 GPU (SP=4, DP=2), lr 1e-5 cosine, max_length 133120, 680 steps; source data/sft/swift_ext_acc/v6-20260825-153450.
• base-acc-step680 — GBS 16, SP 4, micro 1, accum 16, lr 1e-5 cosine, max_length 133120, 680 steps; source data/sft/swift_ext_acc_base/v1-20260903-163953.
• base-longtrace-step680 — same recipe as Base ACC; trained independently of ACC (not continued from it); resumed from checkpoint-300 after a 24h session abort. Best eval during the run was checkpoint-200 (0.129); this subfolder is the aligned step-680 export, not the best-validation checkpoint.

All three omit optimizer/RNG state and the training_args.bin pickle.