specforge
SGLang-EAGLE3-Qwen3-30B-A3B-Instruct-2507-SpecForge-NexSGLang-EAGLE3-Llama-3.1-8B-Instruct-SpecForgeSGLang-EAGLE3-Qwen3-0.6B-SpecForgeSGLang-EAGLE3-Qwen3-Coder-30B-A3B-Instruct-SpecForgeSGLang-EAGLE3-Qwen3-235B-A22B-Instruct-2507-SpecForge-MeituanSGLang-EAGLE3-Llama-3.3-70B-Instruct-SpecForgeSGLang-EAGLE3-Qwen3-Coder-480B-A35B-Instruct-SpecForge-EigenAIQwen3.5-35B-A3B-Eagle3-Specforge
specforge
specforge — a verification benchmark that cannot be memorised
600 protocol-shaped state machines whose ground truth was computed by an exhaustive model checker,
not written down by hand.
Every fixed benchmark has a shelf life: once its answers are in a training corpus, a high score stops
telling you whether a model reasons or remembers. This snapshot is generated, and the generator is
public — so when this set ages, you make a new one with a different seed rather than trusting a… See the full description on the dataset page: https://huggingface.co/datasets/nickh007/specforge.brian-rollouts-311-specforge-turn-conversations
brian-rollouts-311-specforge-turn-conversations
Turn-level conversations SpecForge training export derived from
aimosprite/brian-rollouts-311-compact-turns.
Contents
brian-rollouts-311-specforge-turn-conversations.jsonl
brian-rollouts-311-specforge-turn-conversations.jsonl.manifest.json
Each JSONL row is a single assistant call rendered as a full SpecForge
conversations sample. The visible prefix context is intentionally repeated
across rows so that every assistant turn… See the full description on the dataset page: https://huggingface.co/datasets/aimosprite/brian-rollouts-311-specforge-turn-conversations.brian-rollouts-311-specforge-preformatted-turns
brian-rollouts-311-specforge-preformatted-turns
Preformatted per-turn SpecForge training export derived from
aimosprite/brian-rollouts-311-compact-turns.
Contents
brian-rollouts-311-specforge-preformatted-turns.jsonl
brian-rollouts-311-specforge-preformatted-turns.jsonl.manifest.json
Each JSONL row is a single assistant turn rendered as a full GPT-OSS / Harmony
prompt-plus-completion string in a text field:
{
"id": "amobench::amo-bench-1::attempt0::turn0",
"text":… See the full description on the dataset page: https://huggingface.co/datasets/aimosprite/brian-rollouts-311-specforge-preformatted-turns.brian-rollouts-311-specforge-conversations
brian-rollouts-311-specforge-conversations
Full-rollout conversations SpecForge training export derived from
aimosprite/brian-rollouts-311-compact-turns.
Contents
brian-rollouts-311-specforge-conversations.jsonl
brian-rollouts-311-specforge-conversations.jsonl.manifest.json
Each JSONL row is one reconstructed rollout attempt in SpecForge
conversations format. Harmony messages are reconstructed from the compact
turn snapshots and merged so that repeated assistant text is… See the full description on the dataset page: https://huggingface.co/datasets/aimosprite/brian-rollouts-311-specforge-conversations.sarvam30b-specforge-100k
Sarvam-30B SpecForge Training Data (100k)
Training data for EAGLE3 speculative decoding draft model for sarvamai/sarvam-30b.
Dataset Composition
Source
Samples
mlabonne/open-perfectblend (English/general)
40,000
Hindi (SandLogicTechnologies/Indic_Chat_Dataset)
15,000
Kannada
15,000
Tamil
15,000
Other 10 Indic languages (1,500 each)
15,000
Total
100,000
Format
JSONL with SpecForge conversation format:
{"id": "...", "conversations":… See the full description on the dataset page: https://huggingface.co/datasets/adarshxs/sarvam30b-specforge-100k.specforge_data