license: other
task_categories:
- other
tags: - llm-serving
- scheduling
- conformal-prediction
- activations
TRAIL 14B activations
Per-decode-step hidden activations from deepseek-ai/DeepSeek-R1-Distill-Qwen-14B, layer 27,
collected through vLLM at temperature 0 (greedy). Used to train and calibrate remaining-length
predictors for conformal SRPT scheduling.
Contents
| dir | traces | prompts from |
|---|---|---|
forelen/ |
14,997 | ForeLen pool: math, gsm8k, mmlu-stem, mbpp, cruxeval, ifeval, livecodebench, longseq |
chat20k/ |
20,000 | chat (WildChat / LMSYS derived) |
Each trace_<i>.pt is a torch dict:
| key | |
|---|---|
L27 |
(T, 5120) float16 — last-token hidden state at each decode step |
T |
reasoning length (tokens before </think>) |
g |
(T,) decode step index for each row |
dataset |
source dataset name |
prompt |
the prompt text |
<i> is the prompt's index in the matching shard file, so traces align with the prompt list.
Collection
Activations are captured by a forward hook on layer 27 inside vLLM's model runner, recording the
last-token hidden state on each decode step (prefill steps excluded). Collecting through vLLM
rather than HF generate matters: the predictor is read on the serving path, so training and
serving see the same batched/paged/fp16 distribution.
Traces where </think> appears within 8 tokens, or with fewer than 8 recorded steps, were dropped.
Traces that hit the 16384-token cap are present and record T = 16384, which is not the true
reasoning length — filter them out for training or calibration.
Note on prompts
chat20k prompts derive from WildChat-1M and LMSYS-Chat-1M, which are gated on the Hub with their
own usage agreements. Use of the prompt text here is subject to those upstream terms.