---
license: apache-2.0
language:
- en
task_categories:
- text-generation
tags:
- tool-use
- trajectories
- supervised-fine-tuning
configs:
- config_name: semantic_candidates
  data_files:
  - split: train
    path: data/*.jsonl.gz
---

# Trajectory2Gym: decision-grounded candidate interactions

This release contains **34,772 complete, policy-visible candidate interactions** for tool-use post-training, stored in seven compressed JSONL shards. These are semantic parents, not rendered children or a final scale-specific training view.

| Source | Interactions |
|---|---:|
| bfcl | 17282 |
| tau1 | 11411 |
| tau2 | 6079 |

## Model checkpoints

Download the five corresponding Trajectory2Gym paper checkpoints:

| Scale | Model repository |
|---|---|
| 1.7B | [Trajectory2Gym-1.7B](https://huggingface.co/mis2/Trajectory2Gym-1.7B) |
| 4B | [Trajectory2Gym-4B](https://huggingface.co/mis2/Trajectory2Gym-4B) |
| 8B | [Trajectory2Gym-8B](https://huggingface.co/mis2/Trajectory2Gym-8B) |
| 14B | [Trajectory2Gym-14B](https://huggingface.co/mis2/Trajectory2Gym-14B) |
| 30B-A3B | [Trajectory2Gym-30B-A3B](https://huggingface.co/mis2/Trajectory2Gym-30B-A3B) |

Each model repository contains native tokenizers, chat templates, model cards, licenses, and release checksums. The candidate corpus below is not the rendered scale-specific training data; the 30B-A3B model uses an expanded corpus.

## Record format

Records preserve native messages, tool definitions, supervised assistant-turn indices, context-turn counts, and stable source identities. The supervision mask selects assistant decisions; user/tool messages and masked failed actions can remain visible context.

Operational source paths and non-allowlisted top-level collection metadata are omitted. Dialogue content, native tool arguments, tool definitions, and supervision indices were checked against every original record, including nested dictionary ordering.

## Loading

Native tool schemas and argument values are heterogeneous. Use a JSONL reader rather than assuming all records can be inferred as one flat Arrow schema.

~~~python
import gzip, json
from huggingface_hub import hf_hub_download

path = hf_hub_download('mis2/Trajectory2gym', 'data/train-00000.jsonl.gz', repo_type='dataset')
with gzip.open(path, 'rt', encoding='utf-8') as f:
    for line in f:
        interaction = json.loads(line)
        # interaction['messages'], interaction['supervised_assistant_turn_indices']
~~~

## Scope and provenance

The source is the frozen training-side candidate corpus. Tau1 Airline uses the disclosed seen-task construction/evaluation scope. Tau2 construction uses its official train partition. This release does not assert template-level separation, and it is not the current upstream benchmark test distribution. The 30B-A3B study uses an expanded corpus, not this common pool alone.

Source input SHA-256: a65a690aeb21a8e28c1865607dff93a10cbb1c311c5a7beacf0e58ed3eb94e23

See provenance.json and RELEASE_MANIFEST.json for the transformation and file hashes. Exported bytes differ from the historical source file. Native reset states and all execution certificates are not included, so these files are not a standalone environment replay package.

## Content checks

Credential patterns, internal-host strings, and operational paths were checked. Benchmark-native simulated address literals and a generic path placeholder were retained through an exact-value-hash review bound to the source hash; see NATIVE_LITERAL_REVIEW.json. That review is not an independent human PII audit. Native fixture entities are retained to preserve observation-to-action dependencies.

## License and attribution

Released under Apache-2.0 with the upstream MIT and Apache-2.0 notices preserved. See LICENSE, THIRD_PARTY_NOTICES.md, and licenses/.