leonli66/stage3-final-mixture-20260910-packed

Stage 3 final packed training mixture

Version	Packed sequences	Original examples retained	Parquet files
16,384	2,278,921	21,463,686	4,452
32,768	1,277,010	21,669,429	2,495

Each sequence contains only one category: agent, reasoning or other. Completed sequences are globally shuffled together using PCG64 seed 20260910; both native and expanded agents are included. All shuffle positions and serialized payload preservation were checked. Original examples overlap between versions: choose one version, and do not concatenate them as independent examples.

Lengths count decoder positions after 16:1 compression, including memory boundary tags. Decoder IDs and assistant-only labels are precomputed. Memory strings are encoder-tokenized at runtime; memory vectors and boundary tags are masked from prediction loss. Payloads use the LCLM LargeBinary Arrow format and require its DynamicPackedDataset loader; they are not ordinary text SFT rows.

from huggingface_hub import snapshot_download
root = snapshot_download(
    "leonli66/stage3-final-mixture-20260910-packed", repo_type="dataset",
    allow_patterns=["packed-cs16-32768/**", "release.json", "UPLOAD_STATUS.json"])
dataset_path = root + "/packed-cs16-32768"

Set LCLM stages[3].dataset to dataset_path, max_packed_length to 32768, training.compression_ratio to 16 and data.packed_attention_backend to flash. For 16k, select packed-cs16-16384 and length 16384 instead. Point the loader at the version root containing data/mixed/; retain file and row shuffling and equal work across ranks. Pin the returned Hub revision for reproducible runs.

The tested LCLM build passed single-node H200:8 FSDP and DeepSpeed ZeRO-2 training steps with both lengths. Multi-node/RDMA and full model/optimizer checkpoint restore are not established by those checks. Use ZeRO-2 at most for this build. The loader/source/rank audit records are included under validation/.

Raw input repository.

This is the September 10 final training build, publicly available under leonli66. Upstream terms still apply; this mixture does not grant a replacement license. Preserved upstream notices and source revisions are under source-notices/ and provenance/. Original build manifests retain their Modal paths as provenance. release.json maps the uploaded files to their exact build SHA256 hashes. Check UPLOAD_STATUS.json for verified upload completion.

The raw union has 21,722,787 examples: 20,326,114 base examples, 1,244,170 native agent traces and 152,503 reviewed expanded agent traces. Base reasoning uses the corrected CoT50 policy: prompts remain intact and compression is applied to the selected reasoning text. Native agents remain uncompressed. Expanded agents use memory segments for simulated long tool observations. Rejected and held trace pools are excluded.

Native agents originate from open-thoughts/OpenThoughts-Agent-SFT-100K, nvidia/Nemotron-Agentic-v1 and nvidia/Nemotron-SFT-Agentic-v2; pinned revisions and component counts are in provenance/agent-source-manifest.json and provenance/native-agent-report.json. Expansion selection, source counts and teacher streams are in provenance/expansion-reviewed-transport-manifest.json. The selected expansion union retains reviewed legacy/pilot traces alongside the new Qwen non-thinking traces; it is not exclusively generated by one model.

Decoder tokenizer: Qwen/Qwen3-4B-Instruct-2507, revision cdbee75f17c01a7cc42f958dc650907174af0554. Encoder tokenizer: Qwen/Qwen3-Embedding-0.6B, revision 97b0c614be4d77ee51c0cef4e5f07c00f9eb65b3.

Training code: LeonLixyz/LCLM, train branch. Read the root AGENTS.md for configuration, launch commands and validation scope.