--- pretty_name: CS 329X Delegation Adherence — Blinded Annotation v2 tags: - human-ai-collaboration - delegation - human-annotation - evaluation - coding-agents - outcome-stratified configs: - config_name: collabskill-v2 data_files: - split: annotation path: collabskill-v2/sessions.parquet - config_name: swechat-v2 data_files: - split: annotation path: swechat-v2/sessions.parquet --- # CS 329X Delegation Adherence Dataset ## Current release: blinded annotation v2 The v2 release contains **100 existing source sessions: 50 CollabSkill and 50 SWE-chat**, selected for independent human annotation of delegation adherence and task success. Each source contributes 25 single-turn and 25 multi-turn sessions. These are complete released sessions, with every released event retained by index, rather than cropped episodes or generated synthetic examples. Original-log completeness, correctness and referenced external artifacts have not been independently established. The current annotation folders are **[collabskill-v2](collabskill-v2/)** and **[swechat-v2](swechat-v2/)**. Older `collabskill-curated`, `swe_chat_curated` and `midway-annotation-sample` folders remain historical v1 materials. V2 overlaps with those pools, so adding file counts across versions does not give a unique-session count. Use only the v2 folders for this annotation round. **Status:** Selected and mechanically validated candidate traces, with blank human annotation templates. No human judgments, policy-SFT approvals or trained model are supplied. A protected 80/20 development/test allocation exists locally with the curator. The Hub's `annotation` split contains all 50 sessions in each source configuration and deliberately does not reveal the experimental train/test assignments. | Source | Single-turn | Multi-turn | Total | Hidden train / test | | --- | ---: | ---: | ---: | ---: | | CollabSkill | 25 | 25 | 50 | 40 / 10 | | SWE-chat | 25 | 25 | 50 | 40 / 10 | | Total | 50 | 50 | 100 | 80 / 20 | ## Annotation files and masking Each v2 folder provides: - `scenarios.jsonl`: one session per line, including the released event sequence after the documented projection and redaction. - `sessions.parquet`: viewer-compatible records with `events_json` containing the same event sequence as a JSON string. This avoids assuming identical nested tool schemas across sources. - `scenario_manifest.csv`: opaque display ID, stable scenario ID, dataset, turn type, human-message count and event count. - `index.html` and individual reading files: escaped, inert reading views. - `annotation_template.csv`: blank independent adherence and task-success judgments, evidence, rationale, assessability, confidence and missing-context fields. - `annotation_schema.json`, `README.md` and `checksums.json`: the proposed annotation extension, source attribution, usage and package integrity. **Source outcome grades, outcome bands, grade explanations, human/agent skill ratings and other source analytical labels are omitted from the annotator records. Experimental split labels, split-group IDs and source user/task/repository/session identifiers are also omitted.** Original source event indices are preserved for evidence citation. Stable `scenario_id` and opaque `example_id` permit later reconciliation through the curator's separate local key. The allowlisted behavioral event fields preserve conversation and tool evidence. Contact/credential patterns are redacted throughout retained fields, and source metadata outside the allowlist is omitted. Exact unchanged source records remain local. No new session was shortened by deleting events. Pattern redaction can alter text and is not comprehensive anonymization. User feedback and the agent's natural claims of success or failure remain evidence in the conversation. A source-code task may also contain grade/score words unrelated to source evaluation. **Annotator protocol:** Work only from v2. Do not look up source outcomes, earlier exports or curator files before submitting independent judgments. Public source data and stable scenario hashes mean masking is a protocol against readily supplied labels, not a guarantee against deliberate re-identification. The score lookup and per-session split key are not published in this repository. Reattach them only after independent annotations are frozen, using a validated one-to-one join on `scenario_id` or `example_id`. ## Selection methodology Seed **3292026**, using joint constrained selection and grouped splitting. Inclusion probabilities and design weights are unknown. This is a coverage-oriented curation set, not a representative sample of task prevalence. **Candidate pools:** CollabSkill uses its full pinned eligible inventory, excluding APEX-Agents-derived sessions. SWE-chat uses a seeded bounded reservoir from its full structurally eligible, scored inventory, stratified by initial intent, turn type and outcome band, supplemented by the prior curated pool and additional scores ≥90. The bounded SWE-chat reservoir contains 1,041 candidates. Conversation text is loaded after metadata selection. Missing/invalid source outcomes are excluded and never treated as zero. **Required constraints:** 50 per source, 25 per turn type, equal category/intent counts between single-turn and multi-turn groups, the outcome/split quotas below, at least one training case from every retained category, no known practice groups in test, and no selected related group crossing train/test. SWE-chat retains the cap of two sessions per known user and repository and includes at least eight sessions scoring ≥90, at least two of them in test. Each category with at least four examples spans at least two outcome bands. **Related groups:** Existing transitive source components connect CollabSkill users/tasks and SWE-chat users/repositories. The group's selected members remain within one split. This does not require selecting every source member of the group. Source grouping, input hashes and the private reconciliation key remain local. **Optimization targets:** Maximize task-type coverage in test, then approximate per-category test proportions, GDPval occupation agreement across turn types, and prior multi-turn length proportions. Category/intent quotas are explicit and may differ slightly from a halved v1 sample. Exact occupation matching and exact outcome balance inside every task type are not guaranteed. **Achieved test coverage:** SWE-chat's test set covers all nine intent labels, including `test` and `Unannotated`. CollabSkill's test set covers all nine GDPval sectors. Its three APEX categories remain training-only. Every retained category/intent occurs in train. With only ten test sessions per source, the holdout cannot establish reliable performance for every task type or task-type/outcome combination. `Unannotated` indicates a missing/conflicting source intent, not a verified task class. ### Task-category distribution Every count below divides equally between single-turn and multi-turn. | SWE-chat initial intent | Sessions | | --- | ---: | | Unannotated | 14 | | Create new code | 8 | | Understand | 8 | | Other | 6 | | Debug | 4 | | Refactor | 4 | | Git | 2 | | Test | 2 | | Connect | 2 | | Total | 50 | | CollabSkill source / category | Sessions | | --- | ---: | | GDPval: Professional, Scientific, and Technical Services | 10 | | GDPval: Government | 8 | | GDPval: Finance and Insurance | 6 | | GDPval: Information | 4 | | GDPval: Manufacturing | 4 | | GDPval: Retail Trade | 4 | | GDPval: Wholesale Trade | 4 | | GDPval: Health Care and Social Assistance | 2 | | GDPval: Real Estate and Rental and Leasing | 2 | | APEX: Consulting | 2 | | APEX: Finance | 2 | | APEX: Medicine | 2 | | Total | 50 | GDPval categories use the saved exact task-ID enrichment. APEX categories use the saved reviewed prompt mappings or explicitly marked inferences. These labels have different provenance and should not be treated as equally verified. SWE-chat intents are the source's automated initial-intent labels, not our human adherence judgments. ### Outcome coverage Curator-defined absolute sampling bands are low = [0,50), medium = [50,80), high = [80,100]. Preserve fractional scores. These are sampling intervals, not a shared validated scale or human ground truth. SWE-chat's source top rubric band starts at 90, so scores 80–89 in this package's high band do not mean maximum success. | Partition | Low | Medium | High | Total | | --- | ---: | ---: | ---: | ---: | | Each source, all | 16 | 18 | 16 | 50 | | Each source, train | 13 | 14 | 13 | 40 | | Each source, test | 3 | 4 | 3 | 10 | | Combined, train | 26 | 28 | 26 | 80 | | Combined, test | 6 | 8 | 6 | 20 | Each source also has 8 low, 9 medium and 8 high sessions within each turn type. Each source's test set includes every outcome band in both turn types. Only aggregate sampling quotas are disclosed here, not an example's band or score. **CollabSkill outcome:** The source's session-level `autograde_overall_score`, a 0–100 automated deliverable grade. The usual criteria cover correctness/accuracy, deliverable completeness, task requirements and technical quality. Some source rubrics use custom criteria. This differs from the Bayesian CollabSkill human/agent skill rating. See [CollabSkill §3.2](https://arxiv.org/html/2606.09833v2#S3.SS2). **SWE-chat outcome:** The source's `session_success`, a released LLM-assigned 0–100 session-success score. The source evaluation incorporates task completion, ending state, efficiency, code/commit quality and user experience. It is not an independently rerun correctness test. It differs from attribution percentages and pushback labels. See [SWE-chat Appendix F.2.5](https://arxiv.org/html/2604.20779v2#A6.SS2.SSS5). Analyze outcome distributions separately by source. Their shared numeric range does not make them interchangeable. SWE-chat's interaction criteria and CollabSkill's task-requirements criterion may overlap with adherence, complicating interpretation of associations. Human annotations are needed to identify good collaboration with poor outcomes and poor collaboration with good outcomes. No such human cross-dimension labels have been assumed during selection. ## Proposed task-success annotation dimension The established practice adherence field remains `full`, `partial` or `none`, with evidence and rationale. A separate proposed task-success field uses: | Score | Anchor | | --- | --- | | 0 | No useful progress, unusable output or a failed primary objective supported by evidence. | | 1 | Limited relevant progress; the primary objective remains unmet. | | 2 | Meaningful partial completion; substantial requirements or material defects remain. | | 3 | Core objective achieved; minor omissions or defects remain. | | 4 | All assessable substantive requirements met, with supporting completion/correctness evidence. | Cite original event indices and available artifacts/tests. An assistant's claim of completion alone does not verify success. Missing context is not failure. If insufficient evidence prevents a defensible judgment, set `outcome_assessability=not_assessable`, leave the score blank and explain the missing context. Other assessability values are `adequate` and `partial`; confidence is `low`, `medium` or `high`. A confirmed failure can still support a score even if other evidence is unavailable. This is a proposed extension, not a completed human evaluation or the original papers' numerical scale. The other new dimensions mentioned in the team discussion are unspecified and have not been invented. ## Intended use and limitations The hidden 80-case train partition supports initial label-based judge/model development with broad outcome coverage. Freeze the 20-case test holdout against fitting, rubric iteration and repeated prompt selection. Small subgroup counts limit uncertainty estimates and conclusions. For policy SFT, review strong human collaboration/adherence, supported task success, context usability and applicable training permissions. High source outcomes alone do not approve a demonstration. Low/medium outcome traces can support scoring, contrastive analyses or evaluation without being automatically treated as behavior to imitate. No policy training is performed by this release. Source sessions retain the available released interaction, including tool activity. External task inputs, repositories, generated files, compacted context and original raw transcripts may be unavailable. Conversation length counts structural conversational human messages, not semantically distinct requests. SWE-chat groups logical message fragments using native identifiers. CollabSkill uses conversational user events, including source-marked setup/help prompts. ## Load v2 ```python from datasets import load_dataset import json data = load_dataset("nyartist/cs329x-project", "swechat-v2", split="annotation") record = data[0] events = json.loads(record["events_json"]) ``` Use `collabskill-v2` for the other source. JSONL is also supplied for standard-library loading. No public train/test files are supplied during the blind annotation round. To inspect reading views, download a complete v2 folder and open `index.html` locally. Repository-hosted HTML may be served as source/download rather than executed. ## Sources, attribution and conditions - Shao et al. (2026), [CollabSkill: Evaluating Human-Agent Collaboration on Real-World Tasks](https://arxiv.org/abs/2606.09833). [Trajectory release](https://huggingface.co/datasets/SALT-NLP/cogym-collabskill-trajectories), declared CC BY 4.0. Pinned revision `8c10dd89b4b83ec5a2662f76a70f70572b90a6d8`. - Baumann et al. (2026), [SWE-chat: Coding Agent Interactions From Real Users in the Wild](https://arxiv.org/abs/2604.20779). [Dataset release](https://huggingface.co/datasets/SALT-NLP/SWE-chat), declared ODC-By and applicable source access/content terms. Pinned revision `202b071f18e03c79df5a7287565ce2ebc5ee7756`. - Acknowledge [GDPval](https://huggingface.co/datasets/openai/gdpval) and [APEX](https://huggingface.co/datasets/mercor/APEX-v1-extended) for relevant task-source analyses. [APEX-Agents](https://huggingface.co/datasets/mercor/apex-agents) prohibits training/fine-tuning/parameter fitting in its stated intended-use conditions. APEX-Agents-derived sessions are excluded from v2, although they remain in historical v1 folders. No single blanket license is asserted for this mixed-source package, and embedded repository/task content can have additional rights. Preserve source attribution and review applicable terms for the intended downstream use. V2 changes are constrained selection, blind metadata projection, pattern redaction, normalized packaging, grouped partitioning and blank annotation/rubric files. Source conversations are not synthetic. ## Reproduction, version history and assistance `data/curation/curate_v2.py` in the team's local project records the curation/export procedure. Reproduction requires pinned raw tables, full saved inventories and metadata mappings, the previous pool and practice manifest, and the original curation helpers. These are not all bundled in this Hub release. Private provenance records exact input hashes, solver gaps, runtime versions, event preservation checks, source mappings, outcomes and split assignments. The source code and documentation alone do not establish a complete standalone reproduction environment. **V1:** 200 candidate sessions, no assigned experimental split, plus six exposed practice copies. Historical files are preserved. **V2:** 100 selected sessions with 80/20 grouped allocation, category/intent matching, low/medium/high outcome coverage, restored SWE-chat testing intent, greater SWE-chat ≥90 coverage and blinded annotation packages. Source grades and split assignments remain local for later reconciliation. Previous card wording and exact source exports are preserved locally. Codex assisted with curation code, constrained sampling, mechanical checks, packaging and documentation. All human judgment fields are blank. No LLM adherence judge, human annotation or model training was run for this release.