Downloads · 30 days
0
IDA-TRAIN-V2-mlperf-squad-gpt2
KissTheHabit/IDA-TRAIN-V2-mlperf-squad-gpt2
IDA-TRAIN-V2-mlperf-squad-gpt2 is a machine learning model from KissTheHabit. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Open-weights model fine-tuned on SQuAD v1.1 via the IDA-TRAIN-V2 native engine on an NVIDIA GB10 (Blackwell, sm121a).
Access
Public
Updated Aug 26, 2026
Repo size
3.8 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors3 GB · 81%
At a glance
- Access
- Public
- Created
- Aug 24, 2026
- Updated
- Aug 26, 2026
- SHA
- 619d91f8
Model card
From the Hugging Face model README
IDA-TRAIN-V2-mlperf-squad-gpt2
Open-weights model fine-tuned on SQuAD v1.1 via the IDA-TRAIN-V2 native engine on an NVIDIA GB10 (Blackwell, sm_121a).
Source model: gpt2
Fine-tune dataset: SQuAD v1.1 — 87,599 training Q/A pairs
Hardware: NVIDIA GB10 (Blackwell, sm_121a)
Engine config: CUDA arch 121a, scalar_flash, legacy_bf16, optimizer lion
Artifacts
hf_source_checkpoint/— original Hugging Face checkpoint and tokenizer used as the starting point.native_checkpoint/— checkpoint converted to the IDA native format before training.native_input/— packed SQuAD token/label/segment binary blocks consumed by the native trainer.run/— full training outputs: status, metrics, logs, final native checkpoint, optimizer state, and (for GPT-2) torch fallback checkpoint.
Model architecture
| field | value |
|---|---|
| architecture | ida_lattice |
| hidden size | 768 |
| layers | 12 |
| heads | 12 |
| intermediate size | 3072 |
| vocab size | 50257 |
Training result
| metric | value |
|---|---|
| phase | success |
| final loss | 0.6001923084259033 |
| wallclock (s) | 1707.8678647880151 |
| micro-steps | 2156 |
| optimizer steps | 539 |
| examples seen | 17248 |
| tokens seen | 17661952 |
| sequence length | 1024 |
| microbatch | 8 |
| grad accumulation | 4 |
| global batch size | 32 |
| base learning rate | 0.0003 |
| throughput | 10341.52 tok/s |
| steps/sec | 1.262 |
| ms/step | 792.15 |
| optimizer steps/sec | 0.316 |
| skipped optimizer steps | None |
| final grad norm | None |
| approx params | 163,037,184 |
| compute capability | 12.1 |
| attention backend | scalar_flash |
| precision profile | legacy_bf16 |
| note | torch_fallback_due_to_native_cublas_internal_error |
See the aggregated MLPerf results dataset: KissTheHabit/IDA-TRAIN-V2-mlperf-squad-gb10-results.
- Created
- Aug 24, 2026
- Updated
- Aug 26, 2026