Datasets:

NagaYu
/

kvine-agent-trajectory-bench

depth float32	log_tokens float32	staleness float32	log_access float32	siblings float32	children float32	revisited int64
1	6.398595	0	1.94591	0	1	1
2	3.555348	0	1.609438	0	2	1
3	4.174387	0	0.693147	1	0	1
3	4.60517	0	1.098612	1	0	0
1	6.398595	0	2.302585	0	1	1
2	3.555348	0	2.079442	0	2	1
3	4.174387	0	1.609438	1	1	1
4	4.644391	0	1.386294	0	0	1
3	4.60517	1	1.098612	1	0	0
1	6.398595	0	2.70805	0	1	1
2	3.555348	0	2.564949	0	2	1
3	4.174387	0	2.302585	1	1	1
4	4.644391	0	2.197225	0	1	1
5	3.496508	0	1.609438	0	2	1
6	4.158883	0	0.693147	1	0	1
6	4.59512	0	1.098612	1	0	0
3	4.60517	2	1.098612	1	0	0
1	6.398595	0	2.995732	0	2	1
2	5.777652	0	1.098612	1	0	1
2	3.555348	0	2.772589	1	2	1
3	4.174387	0	2.564949	1	1	1
4	4.644391	0	2.484907	0	1	1
5	3.496508	0	2.079442	0	2	1
6	4.158883	0	1.609438	1	1	1
7	4.60517	0	1.386294	0	0	1
6	4.59512	1	1.098612	1	0	0
3	4.60517	3	1.098612	1	0	0
1	6.398595	0	3.218876	0	3	1
2	5.78996	0	1.098612	2	0	1
2	5.777652	1	1.098612	2	0	1
2	3.555348	0	2.944439	2	2	1
3	4.174387	0	2.772589	1	1	1
4	4.644391	0	2.70805	0	1	1
5	3.496508	0	2.397895	0	2	1
6	4.158883	0	2.079442	1	1	1
7	4.60517	0	1.94591	0	1	1
8	4.624973	0	1.386294	0	0	1
6	4.59512	2	1.098612	1	0	0
3	4.60517	4	1.098612	1	0	0
1	6.398595	0	3.433987	0	4	1
2	5.796058	0	1.098612	3	0	1
2	5.78996	1	1.098612	3	0	1
2	5.777652	0	1.386294	3	0	1
2	3.555348	0	3.091043	3	2	1
3	4.174387	0	2.944439	1	1	1
4	4.644391	0	2.890372	0	1	1
5	3.496508	0	2.639057	0	2	1
6	4.158883	0	2.397895	1	1	1
7	4.60517	0	2.302585	0	1	1
8	4.624973	0	1.94591	0	1	1
9	4.584968	0	1.386294	0	0	1
6	4.59512	3	1.098612	1	0	0
3	4.60517	5	1.098612	1	0	0
1	6.398595	0	3.583519	0	4	1
2	5.796058	1	1.098612	3	0	1
2	5.78996	2	1.098612	3	0	1
2	5.777652	1	1.386294	3	0	1
2	3.555348	0	3.295837	3	2	1
3	4.174387	0	3.178054	1	1	1
4	4.644391	0	3.135494	0	1	1
5	3.496508	0	2.944439	0	2	1
6	4.158883	0	2.772589	1	2	1
7	3.713572	0	1.098612	1	1	1
8	4.60517	0	1.098612	0	1	1
9	4.624973	0	1.098612	0	1	1
10	4.584968	0	1.098612	0	1	1
11	4.65396	0	1.386294	0	0	1
7	4.60517	0	2.564949	1	1	0
8	4.624973	0	2.302585	0	1	0
9	4.584968	0	1.94591	0	1	0
10	3.637586	0	1.098612	0	0	0
6	4.59512	4	1.098612	1	0	0
3	4.60517	6	1.098612	1	0	0
1	6.398595	0	3.78419	0	4	1
2	5.796058	0	1.386294	3	0	0
2	5.78996	0	1.386294	3	0	0
2	5.777652	0	1.609438	3	0	0
2	3.555348	0	3.465736	3	2	1
3	4.174387	0	3.367296	1	1	1
4	4.644391	0	3.332205	0	1	1
5	3.496508	0	3.178054	0	2	1
6	4.158883	0	3.044523	1	2	1
7	3.713572	0	2.079442	1	1	1
8	4.60517	0	2.079442	0	1	1
9	4.624973	0	2.079442	0	2	1
10	3.663562	0	1.098612	1	1	1
11	4.584968	0	1.098612	0	1	1
12	4.65396	0	1.098612	0	1	1
13	4.61512	0	1.386294	0	0	1
10	4.584968	0	1.791759	1	1	0
11	4.65396	0	1.94591	0	1	0
12	3.583519	0	1.098612	0	0	0
7	4.60517	1	2.564949	1	1	0
8	4.624973	1	2.302585	0	1	0
9	4.584968	1	1.94591	0	1	0
10	3.637586	1	1.098612	0	0	0
6	4.59512	5	1.098612	1	0	0
3	4.60517	7	1.098612	1	0	0
1	6.398595	0	3.89182	0	4	1
2	5.796058	1	1.386294	3	0	0

End of preview. Expand in Data Studio

KVine agent-trajectory benchmark & resident-set policy training data 🌿

Two things from the KVine research prototype (branch-aware KV cache for agent trajectories):

policy_train — self-supervised training data for the resident-set policy: structural features of branches in KVine's trajectory tree, labelled with whether the branch was revisited within the next 5 steps.
benchmark_results.json — the full A / B / C benchmark output (cumulative prefill FLOPs, TTFT, recomputed tokens, memory, quality KL, and the policy comparison) on Qwen/Qwen2.5-0.5B-Instruct.

Code / write-up: https://github.com/NagaYu/kvine
Model trained on this data: NagaYu/kvine-resident-policy
Interactive demo: NagaYu/kvine-demo

`policy_train` (`data/policy_train.parquet` / `.csv`)

10,544 rows, positive rate 39.4%, from 12 deterministic seeds of the synthetic agent workload (131 branch revisits total). Each row is one branch snapshot at one step.

column	type	meaning
`depth`	float	depth from root (0 = root)
`log_tokens`	float	`log1p(branch token count)`
`staleness`	float	steps since last access
`log_access`	float	`log1p(access count)`
`siblings`	float	number of sibling nodes
`children`	float	number of child nodes
`revisited`	int (0/1)	label: accessed again within the next 5 steps

The features are structural — they come from the trajectory tree and the workload's access pattern, not from the LLM's weights — so the data is a clean, model-independent signal about which branches of an agent run get revisited.

from datasets import load_dataset
ds = load_dataset("NagaYu/kvine-agent-trajectory-bench", "policy_train")["train"]
print(ds[0])

How it was generated

A pseudo-agent (long system prompt + tool schema as root context) is replayed with stochastic retries / parallel branches / subagents / mid-trajectory observation inserts / branch revisits. Subagent branches (shallow, forked at root) are revisited with higher probability, creating a learnable correlation between the depth feature and revisits. See benchmarks/workload.py in the repo; generation is deterministic per seed.

`benchmark_results.json`

The A / B / C comparison (Naive / Prefix-only=RadixAttention / KVine) at root 8k, 20 steps, plus the 5-seed resident-policy comparison (random 60% → heuristic 81% → learned MLP 87% branch-revisit hit rate).