Fruit Phase-1 training archive
Model-only weights, resumable optimizer/RNG checkpoints, metrics, and plots for GLM-5.2-SIQ-Fruit. This is a 151.54 GB (141.14 GiB) training archive, not a Transformers or vLLM serving repository.
Phase 1 completed 57,293 steps and about 5.43B sampled tokens across MAIN,
LONG, DISTILL, QNOISE, and SFT. The plot is generated from the durable metrics
ledger; raw validation points are in progress/fruit_v1_val.jsonl.
Model-only stage weights
| file | stage | size |
|---|---|---|
final/fruit_v1_main.pt |
MAIN, 4,096 context | 10.081 GB |
final/fruit_v1_long.pt |
LONG, 16,384 context | 10.081 GB |
final/fruit_v1_final.pt |
DSA indexer DISTILL; pre-anneal control | 10.081 GB |
final/fruit_v1_annealed.pt |
QNOISE anneal; published base-model source | 10.081 GB |
final/fruit_v1_instruct.pt |
assistant-masked SFT | 10.081 GB |
The canonical base release source is final/fruit_v1_annealed.pt, SHA-256
98ac7cb4f7799194424782b505d622069fecf4dbca5f5acb2658f2a66c3631f6.
The Instruct source is final/fruit_v1_instruct.pt, SHA-256
32dbf82d40b88a92b8dccd563c593b5971be358cf11895eb150f18644ff93c27.
Resumable checkpoints
checkpoints/*_ckpt.pt and final/fruit_v1_instruct_ckpt.pt are roughly
20.2 GB each. They contain model parameters, optimizer state, step and token
counters, and Python/NumPy/Torch/CUDA RNG state. Use them only with the
matching trainer and geometry in
proxy-fruit.
These full-state files are Python pickle containers and require trusted
torch.load(..., weights_only=False). Do not load a checkpoint from an
untrusted revision. MANIFEST.sha256 authenticates the published archive
files except the card, Git attributes, and the manifest itself.
Legacy convention contract
The Phase-1 weights predate the fail-closed Run-2 markers
serve_conv_v and rope_theta_trained. They use:
- half-split trainer RoPE with theta 500,000;
- trainer MTP
eh_proj(cat[hidden, embed])input order; - stacked routed-expert tensors (
mlp.w_gate,mlp.w_up,mlp.w_down).
Any export or resume must explicitly select the legacy convention. The maintained exporter applies the RoPE channel permutation and MTP input-half swap. Treating these files as serving-native silently corrupts logits and MTP behavior.
Downloading one artifact
Avoid cloning the complete 151 GB repository when only one stage is needed:
hf download malaiwah/fruit-phase1-ckpt \
final/fruit_v1_annealed.pt \
--local-dir ./fruit-checkpoint
sha256sum ./fruit-checkpoint/final/fruit_v1_annealed.pt
For resume, download the matching *_ckpt.pt plus the tokenizer/data inputs
documented by the trainer. Model-only .pt files cannot resume optimizer or
RNG state.
Lineage and evidence
- Serving exports: base SIQ, Instruct SIQ, and BF16 CPU twin.
- Tokenized inputs: fruit-phase1-shards.
- Trainer, stage drivers, restore scripts, smoke suite, and review ledger: github.com/malaiwah/proxy-fruit.
The files are retained for reproducibility and codec experiments. They are not
packaged for AutoModel.from_pretrained() and should not be presented as a
ready-to-serve model.