Hugging Face's logo

Datasets:
spare-rl
/
trsi-proposer-trajectories

Dataset card Files
xet
 Community
trsi-proposer-trajectories / README.md
simonycl's picture
simonycl
README
a0b59f7 verified
preview code
|
Raw
History Blame Contribute Delete
691 Bytes

TRSI raw proposer trajectories — 4-model comparison

Raw proposer sessions (_proposer.json) from the terminal-RSI proposer ablation, one folder per proposer model. Each _proposer.json holds the full session: steps (thinking / tool[bash,read,write,edit] / text), usage (input/output tokens, duration), prompt, system_prompt, and raw_output (raw transcript). Snapshot; kimi-k3 + qwen3.6-35b are still accumulating.

Layout 

<model>/<pool>/full/<wall_id>/v<N>/output/_proposer.json

models: opus (claude-opus-4.8), sol (gpt-5.6-sol), kimi-k3, qwen3.6-35b. pools: armA-* / armtmax-* (opus/sol), armA-k3-48 / armtmax-k3-48 (kimi), prop-qwen35b-s1..s18 (qwen).