docs/trajectory_schema.md

Canonical trajectory and SFT-view schema

The library keeps one immutable, model-agnostic canonical record per imported attempt. Training examples are projections; they never mutate or truncate the canonical record. JSON objects are UTF-8 and may be stored as .jsonl or .jsonl.gz shards.

Canonical record

Trajectory schema version 1.0 contains:

Field	Meaning
trajectory_id	Stable content-addressed ID for one attempt.
task_id / task_key	Native task ID and cross-source normalized task identity. Alternative attempts share task_key.
source	source_id, direct URL, immutable revision, row/file locator and optional file SHA-256.
task_family, domain, split	Benchmark family, environment domain and the untouched upstream split label. Derived train/evaluation assignment lives in task_splits.jsonl.gz.
events	Causally ordered system, user, assistant and tool/environment events.
tools	Available function definitions when published. Missing definitions are declared, never invented.
artifacts	External image/file references and hashes when present.
provenance	Demonstrator class, producer/scaffold, exact model when supplied and source metadata.
outcome	Independent success, status, reward, termination reason and source metadata. null success is unknown, never failure.
verification	Method, whether the source supplied executable verification, and its evidence. Independent reproduction is a separate evidence flag.
completeness	complete, incomplete or truncated. A final submit call may terminate through an externally recorded outcome without a fabricated tool result.
missing_fields	Explicit omissions such as unavailable tool definitions or a missing intermediate result.
transformations	Ordered, versioned adapter operations.
raw_record	Lossless original row/document. Non-JSON bytes are base64 tagged.
metadata	Environment/version, trial, timing, cost and other non-causal attributes.

Each event has a stable event_id, role, text content, zero or more structured tool calls, an optional tool-result link, source metadata, missing fields and a supervised flag. Only assistant events may be supervised. Tool results must refer to an earlier call. Call IDs reused by an upstream source are namespaced deterministically while retaining the original ID.

Canonical validation enforces non-empty identifiers, unique event/call IDs, causal tool-result linkage, assistant-only supervision, declared omissions for incomplete records and preservation of the raw source. It does not manufacture observations or reinterpret a missing outcome as failure.

Independent quality and ranking record

rankings.jsonl.gz is a derived index, kept separate so importing another attempt does not alter canonical records. Every row has separate fields for:

• demonstration_provenance: expert/reference demonstration or recorded agent completion;
• outcome_category: success, failure or unknown;
• verification_strength: independently reproduced, source evaluator, source reported or unverified;
• completeness, source_confidence, quality_tier, numeric tie-break score and within_task_rank;
• SFT eligibility, an interaction hash, duplicate provenance and human-readable ranking reasons.

The five tiers are: independently reproduced successful expert; independently reproduced successful agent; complete demonstration/rollout with credible source evidence; complete rollout with unknown outcome; failed, partial or inadequately observed. “Expert” does not mean optimal. No imported trajectory is called optimal.

Ranks compare only identical task_key plus compatible environment_key. Tier dominates; action-observation coverage, verdict reliability, source confidence and (only for comparable successful attempts) a small efficiency tie-break follow. Identical interactions receive one interaction hash and list all contributing trajectory IDs.

The final library collapses records only when causal interaction, outcome, reward, completeness and verification evidence agree. The retained ranking row lists every contributing ID; duplicate_provenance.jsonl.gz preserves each collapsed source revision and locator. Action-identical rows with conflicting outcome evidence remain separate.

Full-trajectory SFT view

full_trajectory(record) emits one sample:

{
  "view": "full-trajectory",
  "sample_id": "full:...",
  "trajectory_id": "trajectory:...",
  "task_id": "...",
  "task_key": "...",
  "events": ["complete causal event sequence"],
  "target_event_ids": ["every supervised assistant event ID"],
  "model_input": {
    "events": ["complete causal event sequence"],
    "tool_definitions": ["source-published tools"]
  },
  "loss_mask": [false, false, true, false, true],
  "supervised_targets": ["assistant events selected by the mask"],
  "metadata": {"provenance and outcome for filtering/scoring, not model input"}
}

System/user text and returned observations provide context. Loss applies only to assistant events whose mask value is true. A downstream tokenizer maps that event mask to token positions; this library does not impose a tokenizer, chat template, truncation or packing policy.

Assistant-turn SFT view

assistant_turns(record) yields one sample per supervised assistant decision:

{
  "view": "assistant-turn",
  "sample_id": "turn:...",
  "trajectory_id": "trajectory:...",
  "target_event_id": "event:...",
  "target_index": 4,
  "context": ["events 0 through 3 only"],
  "target": {"the recorded assistant event at index 4"},
  "model_input": {
    "events": ["events 0 through 3 only"],
    "tool_definitions": ["source-published tools"]
  },
  "supervised_target": {"the recorded assistant event at index 4"},
  "metadata": {"parent/provenance/outcome; excluded from model_input"}
}

For an assistant event at index i, context is exactly events[:i]. It cannot contain the target, its returned observation, later decisions, outcomes, reference answers or evaluator explanations. Tests assert this invariant. Consumers may shorten or mask the prefix after projection; canonical events remain untouched.

Real example

The successful SWE-Gym attempt trajectory:10b50eb527ccd481a86cc00f targets swe:python__mypy-10308 and has 27 events / 12 assistant decisions:

1. OpenHands system instructions.
2. User task and repository context.
3. Assistant tool call.
4. Returned repository listing.
5. Next assistant tool call.
6. Returned observation, continuing until submission and the source test verdict.

Its full view has one record, 27 event-level mask values and 12 targets. Its turn view has 12 records: turn zero sees events [:2]; turn one sees the first tool result as well; no turn sees a future result.

Concrete outcome cases in the canonical shards are:

• success: trajectory:10b50eb527ccd481a86cc00f (SWE-Gym execution result);
• failure: trajectory:d16d34d7b0acbbd878923aba (same source, failed tests);
• unknown: trajectory:bcf4161ed7c363f7e3e03e1c (TerminalTraj has action/output pairs but no task verdict).

Render any of them with scripts/inspect_library.py --trajectory-id .... The deterministic gallery independently selects compact representatives for every available source/outcome/tier combination and exposes each full interaction, every causal turn sample, the exact model-input object, supervised target, verification evidence, missing fields and side-by-side alternatives.

Consumer selections

dataset.alternatives("swe:python__mypy-10308")  # task-ranked alternatives
dataset.best_verified("swe:python__mypy-10308")
dataset.by_quality_tier(1, 2, 3)

For the large persisted library, use scripts/export_sft_views.py with --selection best, successful or all, plus optional benchmark, source, source-version, model, scaffold, split, quality-tier and quality-rank filters. Deterministic random, task-level and source-balanced sampling use the same CanonicalDatasetAdapter contract documented in dataset_adapter.md. Expert-only means filtering demonstration_provenance in the ranking index; successful-completion-only means outcome_category=success plus the recorded-agent provenance. task_splits.jsonl.gz assigns one split per task_key across every source and both derived views.

Persistent library layout

The default library is manifest-backed: manifest.json lists every lossless pinned canonical shard with its path, SHA-256 and record count. Passing --materialize-all can additionally create all.jsonl.gz. To avoid duplicating very large raw records, outcome and eligibility partitions are compact trajectory-ID indexes:

• successful.index.jsonl.gz, unsuccessful.index.jsonl.gz, and unknown.index.jsonl.gz;
• structurally_eligible.index.jsonl.gz for records passing the causal action-observation gate;
• training_eligible.index.jsonl.gz for structurally suitable records whose source is not protected/evaluation-only.

index.sqlite joins those concepts without loading canonical JSONL and adds source, model, scaffold, task, rank, outcome and quality fields. rankings.jsonl.gz and task_splits.jsonl.gz are portable derived indexes. duplicate_provenance.jsonl.gz records exact rows collapsed into each retained canonical interaction. Partition membership never changes or copies canonical content.