orestis-z/Inkling-Small-NVFP4-Regenerated-Collection
Inkling-Small-NVFP4 Regenerated Collection On-policy training data for a DSpark speculative-decoding drafter targeting thinkingmachines/Inkling-Small-NVFP4. Every assistant response here was regenerated by Inkling-Small-NVFP4 itself over prompts drawn from Magpie + UltraChat, so the completions reflect the target model's own distribution rather than the datasets' original responses. This is what makes the data on-policy for drafter training: the drafter learns to predict the… See the full description on the dataset page: https://huggingface.co/datasets/orestis-z/Inkling-Small-NVFP4-Regenerated-Collection.
Inkling-Small-NVFP4 Regenerated Collection
On-policy training data for a DSpark speculative-decoding drafter targeting `thinkingmachines/Inkling-Small-NVFP4`.
Every assistant response here was regenerated by Inkling-Small-NVFP4 itself over prompts drawn from Magpie + UltraChat, so the completions reflect the target model's own distribution rather than the datasets' original responses. This is what makes the data on-policy for drafter training: the drafter learns to predict the exact tokens the verifier would produce.
How it was built
Produced with the speculators `response_regeneration` pipeline:
- Target model:
thinkingmachines/Inkling-Small-NVFP4, served locally via vLLM (endpoint: http://127.0.0.1:8000/v1/chat/completions). - Regeneration: turn-by-turn. For multi-turn conversations, each assistant turn is regenerated conditioned on the model's own previously regenerated turns (not the original history), producing fully on-policy trajectories.
- Sampling: thinking model at
Thinking effort level: 0.9; responses include explicit<think>chain-of-thought before the final answer.
Source prompts
~500k source conversations; the collection expands to ~630k rows because the pipeline emits one row per assistant turn (multi-turn UltraChat conversations yield several).
The mix is general chat/instruct — it is not concentrated on hard math/code/agentic tasks.
Format
Each row is a regenerated conversation:
The pre-tokenized training variant additionally carries input_ids, a loss_mask marking generated (assistant) tokens, and seq_len (packed to 8192 tokens).
Intended use
Training / distilling speculative-decoding drafters (DSpark, EAGLE-style, etc.) for Inkling-Small-NVFP4. The resulting drafter: `inference-optimization/dspark-inkling-small`.
Notes
- Responses are model-generated and inherit any biases/errors of Inkling-Small-NVFP4; they are not human-verified.
- License follows that of the upstream source prompt datasets (Magpie, UltraChat) — review those before redistribution.