Warm-started from the published DFlash checkpoint poolside/Laguna-S-2.1-DFlash, converted to DSpark format
Dataset: On-policy regeneration of mlabonne/open-perfectblend with Laguna-S-2.1 (RedHatAI/opb-laguna-s)
Hardware: 8x H200 GPUs (4 for vLLM serving with TP4, 4 for training)
Epochs: 3
Loss: {"ce": 0.1, "tv": 0.9}
Learning rate: 1e-4
Sequence length: 8192 tokens, max 3072 anchors
Markov head: rank 256, vanilla
Confidence head: enabled, with Markov, alpha=1.0

Best checkpoint validation metrics

Metric	Value
val/accept_len	6.780
val/accept_rate	0.634
val/loss	0.206
val/ce_loss	0.359
val/tv_loss	0.078
val/position_1_acc	0.857
val/full_acc	0.665

Evaluation (guidellm acceptance length)

Compared against the baseline poolside/Laguna-S-2.1-DFlash:

Subset	DFlash	DSpark	Delta
HumanEval	2.236	2.624	+17.4%
math_reasoning	3.150	3.346	+6.2%
qa	2.167	2.298	+6.1%
question	2.219	2.487	+12.1%
rag	2.277	2.468	+8.4%
summarization	2.177	2.239	+2.8%
tool_call	2.568	2.647	+3.1%
translation	2.948	2.704	-8.3%
writing	2.250	2.466	+9.6%
Average	2.443	2.586	+5.8%

Usage with vLLM

bash

vllm serve poolside/Laguna-S-2.1 \
  --speculative-config '{"method": "dspark", "model": "orestis-z/Laguna-S-2.1-DSpark", "num_speculative_tokens": 15}'

Files

config.json, config.py, model.safetensors — checkpoint files
train_command.txt — full training command with versions
trackio/train_metrics.csv — per-step training metrics (81k steps)
trackio/val_metrics.json — best checkpoint validation metrics