| { |
| "release": { |
| "model_id": "xor-1.2", |
| "release_date": "2026-09-28", |
| "precision": "bfloat16", |
| "parameters": 35107181936, |
| "weight_shards": 16, |
| "candidate": "F12 a24 (lora35-f12, merged at lora_alpha 24)", |
| "received_archive": "f12-a24-champion-52.23-v0.2.1.zip", |
| "received_archive_sha256": "0ba00678790e57555e842302a4540726b1c8503d7b199350c6a5eb4208c7133e" |
| }, |
| "base_model": { |
| "repository": "Qwen/Qwen3.6-35B-A3B", |
| "revision": "995ad96eacd98c81ed38be0c5b274b04031597b0", |
| "revision_evidence": "All 26 weight shards and tokenizer.json of the merge input were hashed and matched the pinned revision; merges.txt, vocab.json and configuration.json matched the Xor 1.1 release copies of the same pinned files." |
| }, |
| "adapter": { |
| "peft_type": "LORA", |
| "rank": 16, |
| "alpha": 32, |
| "merge_alpha": 24, |
| "merge_scale": 1.5, |
| "dropout": 0.05, |
| "target_tables": 310, |
| "config_sha256": "9fe94e209f1662e12cdad136c62b58b21e655a28c21e986618a76064fed71be3", |
| "weights_sha256": "eb52e577d5f9da0945f3743c3e50ccfc7fb20f3aa10ac6b56774ab76c092c1e0", |
| "peft_version": "0.21.0" |
| }, |
| "merge": { |
| "method": "PEFT merge_and_unload(safe_merge=True) at lora_alpha 24 followed by BF16 save_pretrained, then re-split without changing any tensor into the 16-shard layout of Xor 1.1", |
| "script": "merge_box.py (supplied with the adapter archive), variant a24", |
| "script_sha256": "18ba79692fb89cadb7d42073bd60e9d44feeefc70a3fbff55672d3559adad40f", |
| "environment": { |
| "image": "lmsysorg/sglang@sha256:6bcaa47db52f78ce0d67863b8b2431221b79bc23204a80cad757fa819d00e921", |
| "torch": "2.13.0+cu130", |
| "transformers": "5.12.1", |
| "peft": "0.21.0", |
| "safetensors": "0.8.0", |
| "gpu": "1 x NVIDIA H200 NVL" |
| }, |
| "trainer_merge_match": "The unsplit merge output was byte-identical (both weight shards, index and config) to the trainer's own F12 a24 merge.", |
| "reshard": { |
| "method": "Every tensor copied byte for byte into the shard assigned by the Xor 1.1 index; all 1026 tensors compared equal to the unsplit merge afterwards", |
| "layout_template_index_sha256": "f8448aa8b2fbf3519723b0c2965fc94ed64cafdd463381636a1df351952b3f1a", |
| "script_sha256": "276e342d4377e372149f6fb6ff5cd4c0adfc8952393677d6cf66ad2662acb4f9" |
| }, |
| "merged_index_sha256": "f8448aa8b2fbf3519723b0c2965fc94ed64cafdd463381636a1df351952b3f1a", |
| "base_tokenizer_files_copied_from_pinned_base": [ |
| "merges.txt", |
| "vocab.json", |
| "configuration.json" |
| ] |
| }, |
| "serving": { |
| "sglang_image": "prakhar1611/xor-sglang@sha256:94c48d2a6cc98dc456cf93f723707ea7dd81dddfe1061e823b348d68bbe8158f", |
| "sglang_upstream_image": "lmsysorg/sglang@sha256:6bcaa47db52f78ce0d67863b8b2431221b79bc23204a80cad757fa819d00e921", |
| "sglang_patch_dockerfile_sha256": "e05f93d4537cad3e1837fff5e80f511ab1599cf64827cb87aa0853f3d4837d08", |
| "wrapper_sha256": "a44d9dcde4d2e48b43d6b367dae03a0d4e62edd05e639d1a483481435420ec21", |
| "compose_sha256": "24338aa871af898cec3b5dc1262d3e57926aefc5f9fe07e7f19ef231a118d32d", |
| "source_parent_commit": "fed366a6bc7c0f00b53253a4afc9eb29c811f931", |
| "tensor_parallel_size": 1, |
| "data_parallel_size": 1, |
| "marker_count": 255, |
| "marker_table_sha256": "b31e4bbcbaea107ca52f7edb8662061338b78d0172bbcc61052ffb116a524d97", |
| "temperature_map": { |
| "choice": 1.1, |
| "noul": 1.4, |
| "score": 1.0 |
| }, |
| "release_benchmark": "benchmarks/20260928-xor12-release-jevbench.json", |
| "release_benchmark_sha256": "1143ac7225a80466579a3abaa57da2d919dc3edfa96551e9b9db2cae41bdcc14" |
| }, |
| "known_provenance_gaps": [ |
| "The adapter artifact does not contain trainer_state.json.", |
| "adapter_config.json does not record the base revision; the pinned revision was verified by hashing the base files used by the merge before merging.", |
| "adapter_config.json records lora_alpha 32; the release merges at lora_alpha 24 (scale 1.5), the variant selected by the trainer. The alpha change is applied by merge_box.py at merge time.", |
| "The merge used merge_box.py supplied with the adapter rather than serving/merge_adapter.py; both call PEFT merge_and_unload(safe_merge=True).", |
| "The 19 mtp.* (multi-token prediction) tensors of the base checkpoint are not saved by the transformers model class used for the merge, as in Xor 1.1; MTP speculative decoding is not available." |
| ] |
| } |
|
|
