Datasets:
run_id stringclasses 1
value | fs stringclasses 4
values | index int64 0 59 | query stringlengths 27 253 | gold stringlengths 7 172 | reward float64 0 1 | success bool 2
classes | steps int64 2 10 | outcome stringclasses 3
values | messages listlengths 5 22 | generations listlengths 2 10 | full_seq listlengths 2 2 | prompt_len int64 2 2 | turns listlengths 2 10 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
20260828T190903Z_temp0_seed0 | fs_1 | 0 | Calculate a list of duplicate md5 sum hashes for all the ".java" files in the /testbed directory | md5sum /testbed/*.java | awk '{print $1}' | sort | uniq -d | 0.76 | false | 10 | step_limit | [{"role":"system","content":"You are a helpful assistant solving tasks in a Bash shell.\nYou will be(...TRUNCATED) | [{"step":1,"assistant_text":"THOUGHT: First, I need to find all the \".java\" files in the /testbed (...TRUNCATED) | [
"input_ids",
"attention_mask"
] | 2 | [{"action_text":"THOUGHT: First, I need to find all the \".java\" files in the /testbed directory. T(...TRUNCATED) |
20260828T190903Z_temp0_seed0 | fs_1 | 1 | Calculate md5 sum of the md5 sum of all the sorted files under /testbed/dir2/subdir2 | find /testbed/dir2/subdir2 -type f -print0 | sort -z | xargs -r0 md5sum | md5sum | 0.67 | false | 2 | submitted | [{"role":"system","content":"You are a helpful assistant solving tasks in a Bash shell.\nYou will be(...TRUNCATED) | [{"step":1,"assistant_text":"THOUGHT: We need to calculate the md5 sum of all the files under /testb(...TRUNCATED) | [
"input_ids",
"attention_mask"
] | 2 | [{"action_text":"THOUGHT: We need to calculate the md5 sum of all the files under /testbed/dir2/subd(...TRUNCATED) |
20260828T190903Z_temp0_seed0 | fs_1 | 2 | Calculate the md5 sum of the contents of the sorted list of files "$FILES" | cat $(echo $FILES | sort) | md5sum | 0.67 | false | 2 | submitted | [{"role":"system","content":"You are a helpful assistant solving tasks in a Bash shell.\nYou will be(...TRUNCATED) | [{"step":1,"assistant_text":"THOUGHT: I will use the `md5sum` command to calculate the md5 sum of th(...TRUNCATED) | [
"input_ids",
"attention_mask"
] | 2 | [{"action_text":"THOUGHT: I will use the `md5sum` command to calculate the md5 sum of the contents o(...TRUNCATED) |
20260828T190903Z_temp0_seed0 | fs_1 | 3 | Calculate the md5 sum of the md5 sum of all the files sorted under /testbed/dir2/subdir2 | find testbed/dir2/subdir2 -type f -print0 | sort -z | xargs -r0 md5sum | md5sum | 0.67 | false | 10 | step_limit | [{"role":"system","content":"You are a helpful assistant solving tasks in a Bash shell.\nYou will be(...TRUNCATED) | [{"step":1,"assistant_text":"THOUGHT: We need to calculate the md5 sum of all the files under /testb(...TRUNCATED) | [
"input_ids",
"attention_mask"
] | 2 | [{"action_text":"THOUGHT: We need to calculate the md5 sum of all the files under /testbed/dir2/subd(...TRUNCATED) |
20260828T190903Z_temp0_seed0 | fs_1 | 4 | "Calculate the md5 sum of the sorted list of md5 sums of all \".py\" files under /testbed/dir1/subdi(...TRUNCATED) | find /testbed/dir1/subdir1 -type f -name *.py -exec md5sum {} + | awk '{print $1}' | sort | md5sum | 0.67 | false | 10 | step_limit | [{"role":"system","content":"You are a helpful assistant solving tasks in a Bash shell.\nYou will be(...TRUNCATED) | [{"step":1,"assistant_text":"THOUGHT: First, we need to find all \".py\" files under /testbed/dir1/s(...TRUNCATED) | [
"input_ids",
"attention_mask"
] | 2 | [{"action_text":"THOUGHT: First, we need to find all \".py\" files under /testbed/dir1/subdir1. We c(...TRUNCATED) |
20260828T190903Z_temp0_seed0 | fs_1 | 5 | "Calculate the md5sum of each \".py\" file under /testbed/dir1/subdir1, sort the output, and calcula(...TRUNCATED) | "find /testbed/dir1/subdir1 -type f -name \"*.py\" -exec md5sum {} + | awk '{print $1}' | sort | md5(...TRUNCATED) | 0.67 | false | 10 | step_limit | [{"role":"system","content":"You are a helpful assistant solving tasks in a Bash shell.\nYou will be(...TRUNCATED) | [{"step":1,"assistant_text":"THOUGHT: I will use the `find` command to locate all \".py\" files unde(...TRUNCATED) | [
"input_ids",
"attention_mask"
] | 2 | [{"action_text":"THOUGHT: I will use the `find` command to locate all \".py\" files under /testbed/d(...TRUNCATED) |
20260828T190903Z_temp0_seed0 | fs_1 | 6 | "Calculate the total disk usage for each \".txt\" file on the /testbed directory and prepend the sys(...TRUNCATED) | find /testbed / -iname '*.txt' -exec du -s {} + | sed 's/^/$(hostname): /' | 0.67 | false | 10 | step_limit | [{"role":"system","content":"You are a helpful assistant solving tasks in a Bash shell.\nYou will be(...TRUNCATED) | [{"step":1,"assistant_text":"THOUGHT: First, we need to get the hostname of the system. We can use t(...TRUNCATED) | [
"input_ids",
"attention_mask"
] | 2 | [{"action_text":"THOUGHT: First, we need to get the hostname of the system. We can use the `hostname(...TRUNCATED) |
20260828T190903Z_temp0_seed0 | fs_1 | 7 | Change directory to the directory containing the executable file of command "python" | cd $(which python | xargs dirname) | 1 | true | 10 | step_limit | [{"role":"system","content":"You are a helpful assistant solving tasks in a Bash shell.\nYou will be(...TRUNCATED) | [{"step":1,"assistant_text":"THOUGHT: I can use the `which` command to find the path of the \"python(...TRUNCATED) | [
"input_ids",
"attention_mask"
] | 2 | [{"action_text":"THOUGHT: I can use the `which` command to find the path of the \"python\" command. (...TRUNCATED) |
20260828T190903Z_temp0_seed0 | fs_1 | 8 | "Change permissions for all PHP files under the /testbed directory tree to 755 and print the number (...TRUNCATED) | find /testbed -name "*.php" -exec chmod 755 {} \; -exec /bin/echo {} \; | wc -l | 0.67 | false | 10 | step_limit | [{"role":"system","content":"You are a helpful assistant solving tasks in a Bash shell.\nYou will be(...TRUNCATED) | [{"step":1,"assistant_text":"THOUGHT: First, I will use the find command to locate all PHP files und(...TRUNCATED) | [
"input_ids",
"attention_mask"
] | 2 | [{"action_text":"THOUGHT: First, I will use the find command to locate all PHP files under the /test(...TRUNCATED) |
20260828T190903Z_temp0_seed0 | fs_1 | 9 | Check if current shell is running within a 'screen' process. | pstree --show-parents -p $$ | head -n 1 | sed 's/\(.*\)+.*/\1/' | wc -l | 0.78 | false | 10 | step_limit | [{"role":"system","content":"You are a helpful assistant solving tasks in a Bash shell.\nYou will be(...TRUNCATED) | [{"step":1,"assistant_text":"THOUGHT: I can check the parent process ID of the current shell using t(...TRUNCATED) | [
"input_ids",
"attention_mask"
] | 2 | [{"action_text":"THOUGHT: I can check the parent process ID of the current shell using the `ps` comm(...TRUNCATED) |
End of preview. Expand in Data Studio
DeepSeek-Coder-6.7B — InterCode pass-1 agent trajectories
Pass-1 (trajectory generation) of the agent-interpretability emil pipeline:
deepseek-ai/deepseek-coder-6.7b-instruct driven greedily (temp 0, seed 0) through
InterCode Gym environments via a local vLLM OpenAI endpoint. One action per turn,
≤10 turns, every generated token's selected + top-5 log-probs retained. Reward is
InterCode's own execution-based scorer.
Environments were run without Docker (unprivileged host): bash in an Alpine
rootfs via proot, python against InterCode's own python_server.py as a local
process, sql against a local mariadbd + the bundled Spider dev dump. Execution
semantics match the Docker images; see the repo's nodocker_* backends.
Contents
| Env | run-id | episodes | status | avg reward | solved |
|---|---|---|---|---|---|
| intercode-bash | 20260828T190903Z_temp0_seed0 |
200 / 200 | complete | 0.743 | 48 |
| intercode-python | 20260828T185616Z_temp0_seed0 |
261 / 974 | stopped early | 0.379 | 88 |
| intercode-sql | 20260828T204327Z_temp0_seed0 |
50 / 1034 | stopped early | 0.542 | 26 |
The python and sql runs were stopped before finishing; their metadata.json
status is stopped_early and summary reflects only the collected episodes.
Shard manifests were regenerated so each partial run is self-consistent.
Layout
<env>/<run-id>/01-trajectory/
metadata.json run config + summary
validation.json post-hoc validation report (if present)
trajectories.jsonl one JSON record per episode (bash)
trajectory-shards/ trajectories-<start>-<end>.jsonl (python, sql; 25/shard)
trajectory-shards.json shard manifest + hashes (python, sql)
Record schema (per episode)
run_id,environment,index,query,gold,reward,success,steps,outcome(submitted/step_limit/ …)messages— full chat transcript (system / user / assistant / observation)generations[]— per turn:assistant_text,parsed_action,tokens[](token,bytes,logprob,top_logprobs[]),observation, usage/finishevaluation_info— grader detail (python: per-test results; sql: agent/eval rows)- Enrichment (
add_full_seq.py):full_seq(chat-templated token ids),prompt_len,turns[]— so the records also feed the MICE byte-exact reader.
Reproduce
priyankamary/agent-interpretability, branch srujana:
emil/scripts/01-trajectory/{intercode-bash,intercode-python,intercode-sql}/
with INTERCODE_*_NODOCKER=1 and MODEL_ID=deepseek-ai/deepseek-coder-6.7b-instruct.
- Downloads last month
- -