Code-agent trajectories from five benchmark families, four model families and two agent backends (smolagents and react). This is an extension of the BCB hand-over package dated 2026-10-16. It includes all 213 exported run arms, including effort, protocol, critic, and UProp rerun variants. It is not a selection of only the paper's main configurations. Each episode is one task in one arm; the same task can occur in many arms. Do not treat those repetitions as independent tasks.

Benchmark	Arms	Episodes	Finished cohort	Generations
BigCodeBench-Hard	31	3659	2294	15175
LiveCodeBench-Hard	31	10651	7798	39187
CodeContests	30	1759	1262	6505
SWE-bench Verified	59	6783	6082	19816
Polyglot	62	16759	14555	45227

Models: openai/gpt-oss-120b, openai/gpt-oss-20b, Qwen/Qwen3.6-35B-A3B, Qwen/Qwen3.8-27B. Benchmark configuration IDs are bigcodebench_hard, lcb_hard, codecontests, swebench_verified, and polyglot. SWE data here is Verified. Polyglot language and SWE subset/protocol variants are recorded in each run's public configuration. Counts describe exported artifacts, not benchmark sizes.

Download and load

For analysis with Hugging Face Datasets, use one configuration per benchmark:

import json
from datasets import load_dataset

dataset = load_dataset("Alexander4127/agentic_uq_trajectories", "bcb",
                       split="trajectories", streaming=True)
row = next(iter(dataset))
episode = json.loads(row["trajectory_json"])
print(row["arm"], episode["instance_id"], episode["generations"][0]["code"])

The split name trajectories is an inventory, not a prescribed train/test split. Keep all occurrences of a task in the same partition when evaluating generalization across arms. For the reported finished-cohort setting, filter cohort_finished == True within the intended arm and use the documented analysis protocol; pooling arms changes the experiment.

Parquet has stable, filterable columns: benchmark, arm, model, harness, reasoning_effort, instance_id (strings); in_scored_cohort, cohort_finished, success (nullable booleans); n_generations (integer); trajectory_json (string). The last column contains the complete native episode object, serialized as JSON. Encoding that object explicitly preserves dynamic critic/statistic keys, nullable fields and optional sidecars across historical schema variants, without Arrow guessing incompatible nested types. Parse it with json.loads to get the original BCB-style structure. It does not contain the separate controller-step rows.

For the native JSONL files, controller samples and TemporalBelief inputs:

from huggingface_hub import snapshot_download

snapshot_download("Alexander4127/agentic_uq_trajectories", repo_type="dataset",
                  local_dir="agentic_uq_trajectories",
                  allow_patterns=["benchmarks/bcb/**", "manifest.json", "README.md"])

Or download the combined archives/agentic_uq_trajectories_20261103.tar.gz and extract it with tar -xzf. It contains the native layout, documentation, manifest and packer; Parquet files are available separately in the HF repository.

Native layout and schema

README.md
manifest.json
SHA256SUMS
BCB_FORMAT_REFERENCE.md
package_trajectories.py
data/<benchmark>/episodes.parquet
benchmarks/<benchmark>/arms/<model>/<harness>/<tag>/
    episodes.jsonl
    temporal_belief_episodes.jsonl
    run_config.json
    steps.jsonl                      # only where controller samples exist
archives/agentic_uq_trajectories_20261103.tar.gz

The benchmarks/<benchmark>/ prefix is the only layout extension over the original BCB package. Every arm has its own files and an entry in manifest.json, including model, backend, effort, benchmark, counts and public run arguments. Public configs exclude machine-local output paths, endpoint/network settings and credential-related fields. The original source exports remain unchanged.

episodes.jsonl has one row per candidate-text task, including artifacts outside the compact scored cohort:

Field	Meaning
`instance_id`	Original task ID, unique within this arm
`in_scored_cohort`	A compact analysis record exists
`cohort_finished`	Compact record has `exit_reason == "finish"` and no `hit_cap`
`success`	Final-candidate oracle outcome; null if not exported
`metadata`	Exported step counts, final action, cap and exit reason
`episode_critics`	Last exported verdict of each critic in the episode
`generations`	Ordered exported candidates, with fields below

Each generation contains index, step_index, code (candidate code or SWE patch), code_sha1, label_pass_hidden_tests, label_available, critics, and token_stats. code preserves the exported raw candidate text, which may include Markdown fences; code_sha1 is the original grader's digest of extracted code and therefore need not equal the digest of the raw code string. Critic meanings depend on the environment/protocol; consult the corresponding run and source implementation. Their verdicts are observations, not ground-truth labels. Do not feed oracle labels or final success into uncertainty predictors.

For legacy LCB labels with no available field, label_available is inferred from the presence of a boolean Y. A confirmed field marks relabel provenance, not availability. Missing source labels remain null; unavailable labels remain distinct from failed candidates. No missing result is filled with false.

UProp sidecars, where exported, add generation reasoning_text, answer_text, text_source, text_clipped, finish_reason, prompt_tokens, completion_tokens, and token_confidence_summary; episode htc and n_steps_with_samples. steps.jsonl preserves controller-step records including episode ID, turn/step indices, realized decision, alternative samples with NLL/token counts, parsing counts and elapsed time. These samples are auxiliary observations of the same trajectory, not independent trajectories.

temporal_belief_episodes.jsonl contains only the finished cohort, in the toolkit's original schema: episode_id, harness, success, n_steps_total, and generations[{index, step_index, critics, signals}]. It uses compact generations as the original BCB packer did. Full episodes.jsonl keeps every candidate even where the compact export omitted a generation; its absent statistics remain null. There are 25 episodes with differing candidate/compact generation counts (24 in q36rest_smol_swe, one in uprop_oss20_smol_lcb). There are 122 candidate-only episodes across the five benchmarks. Use the cohort flags rather than assuming all artifacts have a scored analysis record.

In q36rest_smol_swe, resumed exports also contain duplicate source generation indices: the candidate-text cache retained the texts but discarded their original indices. For affected episodes, candidate_alignment is ambiguous_duplicate_source_indices and candidate-level labels, digests, step indices, critics and statistics are null rather than guessed. Episode metadata and compact-derived TemporalBelief rows retain the historical analysis record; they do not establish a candidate-text alignment. Original labels, compact records and the affected instance list are preserved under the arm's alignment_provenance/. The same conservative treatment applies to episode 2839 in uprop_oss20_smol_lcb, whose generation sidecar has a duplicate index; its original generation sidecar is also preserved in alignment_provenance/. The manifest reports the number of affected episodes. Unaffected BCB rows preserve the original package exactly.

Some sidecar records refer to candidates absent from the text cache. Such records do not become invented candidate-text episodes. The manifest counts orphan keys, and alignment_provenance/orphan_keys.json plus the original source sidecar/labels preserve them for inspection. Controller samples remain in steps.jsonl, including steps without a matching candidate-text episode.

Token statistics and unavailable information

token_stats uses three channel prefixes: none for the whole completion, reasoning_ for the thinking channel, and answer_ for the visible answer.

Statistic	Definition
`sum_logprob`	Sum of sampled-token log probabilities
`mean_token_logprob`	Mean sampled-token log probability
`min_token_logprob`	Minimum sampled-token log probability
`num_tokens`	Number of sampled tokens in the channel
`perplexity`	`exp(-mean_token_logprob)`, derived during packaging
`mean_token_entropy`	Mean entropy in nats of the renormalized top-k distribution
`self_certainty`	Mean `-mean_v(log p(v)) - log(k)` from raw top-k probabilities

Higher sum/mean logprob and self-certainty indicate greater confidence; higher perplexity and entropy indicate lower confidence. Top-k was 20 in the campaign; check each public run configuration for its exact parameters. Missing channel statistics remain missing. See BCB_FORMAT_REFERENCE.md for the original glossary and TemporalBelief input description.

Per-token top-1/top-20 arrays are unavailable. They were removed after aggregate export under the storage decision recorded on 2026-09-19, and the source job's temporary storage is gone. This release does not reconstruct them. Prompts, tool observation text and judge-feedback text were not exported. Full reasoning and answer text exists only where sidecars were preserved; candidate text alone is not a full conversational transcript. text_clipped and text_source must be respected when working with sidecars.

Provenance and reproduction

Sources are the campaign's reports/exports/<tag>/ candidate-text, generation-label, compact-v2, run-config, and optional generation/UProp sidecar files. The packer joins on (instance_id, generation_index), detects ambiguous duplicate keys and preserves orphan records separately. It requires all BCB episode, TemporalBelief and controller rows to compare equal to the original 2026-10-16 package. Public run configs are intentionally filtered; trajectory objects preserve the hand-over schema.

Run on an authorized compute job with an existing PyArrow environment:

python scripts/package_trajectories.py \
  --exports /path/to/reports/exports \
  --reference-bcb /path/to/reports/bcb_trajectories_20261016 \
  --output /path/to/reports/agentic_uq_trajectories_20261103

Output must not exist. No source export is modified. SHA256SUMS covers published files and the combined archive (its path is recorded under archives/). The checksum file is outside the tar to avoid a self-referential archive checksum. No model is run and no benchmark code is executed during packaging.

To publish the prepared package, run python scripts/upload_trajectories.py /path/to/reports/agentic_uq_trajectories_20261103. Supply authentication through HF_TOKEN or the hidden terminal prompt. The uploader does not save the token, preserves repository visibility and existing unrelated files, and checks the published file inventory and sizes at the resulting revision.

This release does not assign a new blanket license to third-party benchmark material or model-generated content. Consult the upstream benchmark/model terms and the source project's license before redistribution or downstream use.

Downloads last month: 42

Total file size:

2.82 GB