Datasets:
|
Download README.md from MINZIK77/midtraining-pool: direct link, hf CLI and curl.
- Browser
- Download file 31.5 kB
-
https://huggingface.co/datasets/MINZIK77/midtraining-pool/resolve/main/README.md
- Command line
-
hf download hf://datasets/MINZIK77/midtraining-pool/README.md
-
curl -L -o README.md https://huggingface.co/datasets/MINZIK77/midtraining-pool/resolve/main/README.md
31.5 kB
| license: other | |
| license_name: mixed | |
| license_link: LICENSE.md | |
| task_categories: | |
| - text-generation | |
| language: | |
| - en | |
| tags: | |
| - agent | |
| - code | |
| - software-engineering | |
| - trajectories | |
| - mid-training | |
| size_categories: | |
| - 1M<n<10M | |
| configs: | |
| - config_name: trajectory-nanoswe | |
| data_files: | |
| - split: train | |
| path: data/trajectory/nanoswe/*.parquet | |
| - config_name: trajectory-minicoder | |
| data_files: | |
| - split: train | |
| path: data/trajectory/minicoder/*.parquet | |
| - config_name: trajectory-klear | |
| data_files: | |
| - split: train | |
| path: data/trajectory/klear/*.parquet | |
| - config_name: trajectory-openswe | |
| data_files: | |
| - split: train | |
| path: data/trajectory/openswe/*.parquet | |
| - config_name: trajectory-coderforge | |
| data_files: | |
| - split: train | |
| path: data/trajectory/coderforge/*.parquet | |
| - config_name: trajectory-nebius_oh | |
| data_files: | |
| - split: train | |
| path: data/trajectory/nebius_oh/*.parquet | |
| - config_name: trajectory-swezero | |
| data_files: | |
| - split: train | |
| path: data/trajectory/swezero/*.parquet | |
| - config_name: trajectory-nemoswe_v1 | |
| data_files: | |
| - split: train | |
| path: data/trajectory/nemoswe_v1/*.parquet | |
| - config_name: trajectory-nemoswe_v2_oh | |
| data_files: | |
| - split: train | |
| path: data/trajectory/nemoswe_v2_oh/*.parquet | |
| - config_name: trajectory-scaleswe | |
| data_files: | |
| - split: train | |
| path: data/trajectory/scaleswe/*.parquet | |
| - config_name: trajectory-sera | |
| data_files: | |
| - split: train | |
| path: data/trajectory/sera/*.parquet | |
| - config_name: trajectory-terminal | |
| data_files: | |
| - split: train | |
| path: data/trajectory/terminal/*.parquet | |
| - config_name: trajectory-otagent | |
| data_files: | |
| - split: train | |
| path: data/trajectory/otagent/*.parquet | |
| - config_name: trajectory-agenttrove | |
| data_files: | |
| - split: train | |
| path: data/trajectory/agenttrove/*.parquet | |
| - config_name: trajectory-opencode | |
| data_files: | |
| - split: train | |
| path: data/trajectory/opencode/*.parquet | |
| # Mid-training pool | |
| Public agent datasets in one schema, for agentic mid-training. The pool applies light filters only (see | |
| [Filters](#filters)). A training mix selects and weights rows from it. | |
| This version has one category, `trajectory`: 15 configs, 6,001,371 rows and 154.59B tokens (sigdig | |
| tokenizer). Each config is `trajectory-<source>` and reads `data/trajectory/<source>/`. | |
| ## Categories | |
| | Group | Content | Configs | Rows | Tokens (B, sigdig) | | |
| |---|---|---|---|---| | |
| | SWE, mini-swe-agent and two-tool harnesses | Issue resolution in a code repository; harnesses mini-swe-agent, mini-swe-agent-plus, or the tools bash and file_editor. | `trajectory-nanoswe`, `trajectory-minicoder`, `trajectory-klear` | 1,788,775 | 34.14 | | |
| | SWE, OpenHands and other harnesses | Issue resolution in a code repository; harnesses OpenHands, SWE-agent, mini-swe-agent, opencode, claude-code or not named. | `trajectory-openswe`, `trajectory-coderforge`, `trajectory-nebius_oh`, `trajectory-swezero`, `trajectory-nemoswe_v1`, `trajectory-nemoswe_v2_oh`, `trajectory-scaleswe`, `trajectory-sera` | 1,961,140 | 92.24 | | |
| | Terminal and general agent tasks | Terminal, CLI and other agent tasks (AgentTrove also has code repair, math and competitive programming). | `trajectory-terminal`, `trajectory-otagent`, `trajectory-agenttrove`, `trajectory-opencode` | 2,251,456 | 28.21 | | |
| ## Load | |
| ```python | |
| from datasets import load_dataset | |
| ds = load_dataset("MINZIK77/midtraining-pool", "trajectory-nanoswe", split="train", streaming=True) | |
| row = next(iter(ds)) | |
| ``` | |
| ## Schema | |
| | Column | Content | | |
| |---|---| | |
| | `id` | `source/file/row` (parquet sources) or `source/file@byte_offset` (jsonl sources) in the source dataset | | |
| | `source` | source name (the config without the `trajectory-` prefix) | | |
| | `task_id`, `repo` | task (instance) id and repository, when the source gives them | | |
| | `scaffold` | agent harness as named by the source card, or `not named` | | |
| | `outcome` | `resolved`, `failed`, `incomplete` or `unlabeled`; see [Labels](#labels) | | |
| | `group` | subset of the source (folder, split or source label); see the source section | | |
| | `messages` | turns: `role` (system, user, assistant, tool), `content`, `reasoning`, `tool_calls` (JSON list of id, name, arguments), `tool_call_id`, `name` | | |
| | `tools` | tool definitions of the source, as JSON text (null when the source has none) | | |
| | `n_messages` | number of turns | | |
| | `n_chars`, `n_tokens_sigdig`, `n_tokens_smollm3` | characters and tokens of a ChatML rendering of the row (tools as one system turn; reasoning, content and tool calls of each turn), with [allenai/dolma2-tokenizer-sigdig](https://huggingface.co/allenai/dolma2-tokenizer-sigdig) @86d32e8 and [HuggingFaceTB/SmolLM3-3B](https://huggingface.co/HuggingFaceTB/SmolLM3-3B), no special tokens | | |
| | `content_hash` | blake2b-64 of that rendering, for deduplication across sources | | |
| The rendering is for counting only. The rows keep the source text, so a mix can apply its own chat template. | |
| ## Labels | |
| `resolved` and `failed` come from a success field of the source. `incomplete` means the source records that the run | |
| stopped with an error. `unlabeled` means the source has no such field; it does not mean that the run failed or | |
| succeeded. | |
| | Config | Source field and mapping | | |
| |---|---| | |
| | `trajectory-nanoswe` | `verified`: "this rollout resolved its task" (the teacher campaign's per-rollout grade). The card keeps both resolved and unresolved rollouts in `train`. | | |
| | `trajectory-minicoder` | `verified` (boolean). The card does not define it; the pool reads `true` as resolved and `false` as failed. | | |
| | `trajectory-klear` | No label field; all rows are `unlabeled`. | | |
| | `trajectory-openswe` | `resolved`: 1 if the agent solved the issue, 0 otherwise, -1 if the information is unavailable. The pool maps 1 to resolved, 0 to failed and -1 to unlabeled. | | |
| | `trajectory-coderforge` | `reward` (float). The card does not define it; the pool reads `reward >= 1.0` as resolved and lower values as failed. | | |
| | `trajectory-nebius_oh` | `resolved` (1 solved, 0 not) and `exit_status` (`submit` when the agent ended with its terminating action, else an OpenHands error message). The pool sets resolved if `resolved` is 1, failed if 0 and the run was submitted, and incomplete if 0 and the run ended with an error. | | |
| | `trajectory-swezero` | No label field; all rows are `unlabeled`. | | |
| | `trajectory-nemoswe_v1` | No label field; all rows are `unlabeled`. | | |
| | `trajectory-nemoswe_v2_oh` | No label field; all rows are `unlabeled`. | | |
| | `trajectory-scaleswe` | No label field; all rows are `unlabeled`. | | |
| | `trajectory-sera` | No label field; all rows are `unlabeled`. "Submitted" does not mean resolved. | | |
| | `trajectory-terminal` | No label field; all rows are `unlabeled`. | | |
| | `trajectory-otagent` | No label field; all rows are `unlabeled` (the source's `result` column is described only as rollout bookkeeping). | | |
| | `trajectory-agenttrove` | The card documents a `reward` column (typically 1.0 success, 0.0 failure, missing for some sources). The files at this revision have a `result` column instead; the pool maps `1.0` to resolved, `0.0` to failed, error class names (for example `AgentTimeoutError`) to incomplete and empty values to unlabeled. This mapping is the pool's convention. | | |
| | `trajectory-opencode` | No label field; all rows are `unlabeled`. | | |
| ## Filters | |
| 1. Rows with no assistant turn, or that do not parse, are removed. | |
| 2. Rows of the 12 SWE-bench repositories (astropy, django, matplotlib, seaborn, flask, requests, xarray, pylint, | |
| pytest, scikit-learn, sphinx, sympy) are removed, also forks with the same name. The match uses the repository | |
| name or the task id, so sources without them are not filtered. | |
| 3. Exact duplicates of the rendering inside a source are removed; the first one in file order stays. | |
| Failed, incomplete and unlabeled runs stay. Rollouts of the same task are not capped. | |
| ## Sources | |
| ### SWE, mini-swe-agent and two-tool harnesses | |
| #### `trajectory-nanoswe` | |
| Source: [nanoswe/nanoswe-trajs-261002](https://huggingface.co/datasets/nanoswe/nanoswe-trajs-261002) at `64d22ccf2c` | |
| - One multi-turn SWE agent trajectory per row (issue, then a tool-calling rollout). The first turn is the bare issue text; the teacher's system prompt and tool schema are removed. | |
| - Teacher model: `nemotron-3.5-lightning`, acting through two tools, `bash` and `file_editor`. The card does not name the harness. | |
| - Tasks: SWE-smith-style task environments (campaigns `swe-smith-lightning` and `swe-smith-v3-lightning`). The card states 1,133 repositories, none of them one of the 12 SWE-bench source repositories or a fork of one. | |
| - Source processing (card): a trajectory is cut at the first turn that the format cannot represent; trajectories with no tool call are dropped; observations are capped at 100,000 characters. | |
| - Only the `train` split is in the pool (the card's `validation` split holds 500 held-out task instances). | |
| - Label: `verified`: "this rollout resolved its task" (the teacher campaign's per-rollout grade). The card keeps both resolved and unresolved rollouts in `train`. | |
| - Used for: Training corpus of [nanoswe/nanoswe-192h-261002](https://huggingface.co/nanoswe/nanoswe-192h-261002), a 4.99B SWE agent model trained from random initialization on these trajectories, with the loss on assistant spans. | |
| - License: MIT. Code quoted in tool observations stays under the licenses of the upstream projects. | |
| - Pool note: In the source, assistant turns keep their text in `parts` and leave `content` empty; the pool moves `parts` into `content`, as the card advises for other readers. | |
| - In this pool: 1,328,316 rows, 27.83B tokens (sigdig), 28.06B (SmolLM3); outcome resolved 670,598, failed 657,718. Removed: 55,506 SWE-bench rows, 0 duplicates, 0 with no assistant turn or unparsable. | |
| #### `trajectory-minicoder` | |
| Source: [ricdomolm/mini-coder-trajs-400k](https://huggingface.co/datasets/ricdomolm/mini-coder-trajs-400k) at `90b240e098` | |
| - SWE agent trajectories generated with [Qwen/Qwen3-Coder-30B-A3B-Instruct](https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct) in [mini-swe-agent](https://mini-swe-agent.com/latest/), on SWE-smith tasks. | |
| - Card metadata: 396,515 rows with the columns `messages`, `patch`, `model`, `instance_id` and `verified`. | |
| - Label: `verified` (boolean). The card does not define it; the pool reads `true` as resolved and `false` as failed. | |
| - Used for: The card says it was used to train the [mini-coder](https://huggingface.co/collections/ricdomolm/mini-coder-68dbc8198dd63ff63ed8ef7d) models. The model cards of mini-coder-4b (from Qwen3-4B-Instruct-2507) and mini-coder-1.7b (from Qwen3-1.7B) list this dataset and describe both as distilled from Qwen 3 Coder 30B A3B. | |
| - License: MIT (card metadata). | |
| - In this pool: 394,469 rows, 5.14B tokens (sigdig), 5.14B (SmolLM3); outcome failed 310,516, resolved 83,953. Removed: 2,046 SWE-bench rows, 0 duplicates, 0 with no assistant turn or unparsable. | |
| #### `trajectory-klear` | |
| Source: [Kwai-Klear/SWE-smith-mini_swe_agent_plus-trajectories-66k](https://huggingface.co/datasets/Kwai-Klear/SWE-smith-mini_swe_agent_plus-trajectories-66k) at `750b2c1123` | |
| - About 66k issue-solving trajectories collected with [mini-swe-agent-plus](https://github.com/Kwai-Klear/mini-swe-agent-plus) on issues derived from SWE-smith. Each row has only `instance_id` and `messages`. | |
| - The card does not name the model that generated the trajectories, and does not say whether they are filtered to resolved runs. | |
| - Label: No label field; all rows are `unlabeled`. | |
| - Used for: The card reports training Qwen3-8B on 1k to 66k of these trajectories (SWE-bench Verified solve rate rises about linearly with the log of the data size) and lists Klear-Agent-8B-SFT, trained on this dataset with mini-swe-agent-plus, at 39.0% on SWE-bench Verified. | |
| - License: MIT (card metadata). | |
| - In this pool: 65,990 rows, 1.17B tokens (sigdig), 1.17B (SmolLM3); outcome unlabeled 65,990. Removed: 0 SWE-bench rows, 4 duplicates, 0 with no assistant turn or unparsable. | |
| ### SWE, OpenHands and other harnesses | |
| #### `trajectory-openswe` | |
| Source: [nvidia/Open-SWE-Traces](https://huggingface.co/datasets/nvidia/Open-SWE-Traces) at `4fa4c2ae9a` | |
| - NVIDIA dataset of SWE agent trajectories collected with five harnesses: OpenHands, SWE-agent, mini-swe-agent, opencode and claude-code. The card names models of the MiniMax, Qwen, DeepSeek and GLM families as generators. | |
| - Issue statements and repository environments come only from [Scale-SWE](https://huggingface.co/datasets/AweAI-Team/Scale-SWE) and [SWE-rebench-V2](https://huggingface.co/datasets/nebius/SWE-rebench-V2). | |
| - The pool reads every file of the card's configs v1.0 to v2.0. In the repository, files sit in `data/<harness>/<model>/<scale-swe or swe-rebench-v2>/`, some in a further `thinking-low` or `thinking-medium` folder; the pool's `group` column keeps this path and `scaffold` keeps the harness folder. Release notes on the card: v1.1 adds DeepSeek-V4-Flash-0731 and Qwen3.6-27B, v1.2 adds Qwen3.8-27B (xhigh reasoning effort), v2.0 adds Qwen3.8-Flash-Next at xhigh, medium and low reasoning effort. | |
| - For v1.0, the authors removed trajectories that showed git hacking behavior (card release notes); the card does not say this for later versions. | |
| - Intended use (card): supervised fine-tuning and distillation of SWE agents. | |
| - Label: `resolved`: 1 if the agent solved the issue, 0 otherwise, -1 if the information is unavailable. The pool maps 1 to resolved, 0 to failed and -1 to unlabeled. | |
| - Used for: The card does not name a model trained on it. | |
| - License: CC BY 4.0; ready for commercial and non-commercial use. Each row of the source records the SPDX license of its repository. | |
| - In this pool: 1,119,035 rows, 60.92B tokens (sigdig), 61.27B (SmolLM3); outcome failed 572,620, resolved 483,937, unlabeled 62,478. Removed: 19,910 SWE-bench rows, 0 duplicates, 0 with no assistant turn or unparsable. | |
| #### `trajectory-coderforge` | |
| Source: [togethercomputer/CoderForge-Preview](https://huggingface.co/datasets/togethercomputer/CoderForge-Preview) at `060fca96cf` | |
| - Together AI coding-agent SFT trajectories. The card describes it as a test-verified dataset built from 51K distinct open-source tasks, skewed toward bug fixing. | |
| - All trajectories come from one scaffold with a fixed tool set; the card does not name the scaffold or the model that generated the trajectories. | |
| - The pool reads the splits `SWE_Rebench` (77,169 rows), `SWE_Smith` (148,001) and `R2E_Gym` (32,964); the `group` column keeps the split. The card's fourth split `filtered_reward1` (155,144 rows, in the same folder) is not read: in a first build that read it, 151,795 of its rows were exact copies of rows of the other splits. | |
| - Card limitation: models trained with SFT on this data can do worse with other scaffolds, tools, prompt templates or tool-call formats. | |
| - Label: `reward` (float). The card does not define it; the pool reads `reward >= 1.0` as resolved and lower values as failed. | |
| - Used for: The card reports that fine-tuning Qwen-3 32B on it raised SWE-Bench Verified pass@1 from 23.0% to 59.4%; it does not name a released model. | |
| - License: No dataset license is stated on the card. Rows of the source carry a `license` field that the card does not describe. | |
| - In this pool: 251,676 rows, 10.73B tokens (sigdig), 10.86B (SmolLM3); outcome resolved 151,794, failed 99,882. Removed: 6,457 SWE-bench rows, 1 duplicates, 0 with no assistant turn or unparsable. | |
| #### `trajectory-nebius_oh` | |
| Source: [nebius/SWE-rebench-openhands-trajectories](https://huggingface.co/datasets/nebius/SWE-rebench-openhands-trajectories) at `35455389ab` | |
| - Multi-turn trajectories of [Qwen/Qwen3-Coder-480B-A35B-Instruct](https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct) in OpenHands v0.54.0, on real GitHub issues from [nebius/SWE-rebench](https://huggingface.co/datasets/nebius/SWE-rebench). | |
| - Card totals: 67,074 trajectories, 32,161 of them successful; 1,823 repositories; at most 100 turns, 64.3 on average. | |
| - Both `user` and `tool` turns carry environment observations. Tool-call arguments are stored as strings. | |
| - Label: `resolved` (1 solved, 0 not) and `exit_status` (`submit` when the agent ended with its terminating action, else an OpenHands error message). The pool sets resolved if `resolved` is 1, failed if 0 and the run was submitted, and incomplete if 0 and the run ended with an error. | |
| - Used for: The card does not name a model trained on it. | |
| - License: CC BY 4.0. | |
| - In this pool: 66,008 rows, 3.47B tokens (sigdig), 3.50B (SmolLM3); outcome resolved 31,551, failed 29,328, incomplete 5,129. Removed: 1,066 SWE-bench rows, 0 duplicates, 0 with no assistant turn or unparsable. | |
| #### `trajectory-swezero` | |
| Source: [nvidia/SWE-Zero-openhands-trajectories](https://huggingface.co/datasets/nvidia/SWE-Zero-openhands-trajectories) at `7b3cd106d0` | |
| - 318,115 OpenHands trajectories on 118,092 issues, generated by Qwen3-Coder-480B-A35B-Instruct. The paper calls them execution-free trajectories; the card's title calls the method execution-free fine-tuning. | |
| - Issue statements come from SWE-Gym, SWE-Gym-Raw, R2E-Gym-Subset (issue statements written by Qwen3-Coder-480B-A35B-Instruct from its prompts), SWE-Fixer-Train-110K and SWE-rebench. | |
| - Label: No label field; all rows are `unlabeled`. | |
| - Used for: Stage one (SWE-ZERO) of the paper's two-stage SFT recipe SWE-ZERO to SWE-HERO, which trains agents based on the Qwen2.5-Coder series; the second stage (SWE-HERO) applies execution-backed refinement. The paper reports 62.2% on SWE-bench Verified for SWE-HERO-32B. | |
| - License: CC BY 4.0; ready for commercial and non-commercial use. The card states that only repositories under MIT, Apache-2.0, BSD-2-Clause and BSD-3-Clause are used. | |
| - In this pool: 308,585 rows, 8.09B tokens (sigdig), 8.25B (SmolLM3); outcome unlabeled 308,585. Removed: 9,528 SWE-bench rows, 2 duplicates, 0 with no assistant turn or unparsable. | |
| #### `trajectory-nemoswe_v1` | |
| Source: [nvidia/Nemotron-SWE-v1](https://huggingface.co/datasets/nvidia/Nemotron-SWE-v1) at `0fe17a965b` | |
| - NVIDIA SFT dataset of OpenHands trajectories generated by Qwen3-Coder-480B-A35B-Instruct, curated to improve SWE-Bench-style performance. The card names SWE-Gym and R2E-Gym-Subset as issue sources. | |
| - The pool reads the card's only split, `r2e_gym` (`data/r2e_gym.jsonl`, 51,029 rows). The card text says 59k trajectories; its table lists only this split. | |
| - Label: No label field; all rows are `unlabeled`. | |
| - Used for: The card does not name a model trained on it. It lists the SWE-Gym, R2E-Gym and OpenHands Software Agent SDK papers as references. | |
| - License: CC BY 4.0, except certain subsets under BSD-3-Clause as the Hugging Face Dataset Viewer shows; Apache 2.0 and MIT are listed as additional information. Ready for commercial use. | |
| - In this pool: 51,029 rows, 2.38B tokens (sigdig), 2.40B (SmolLM3); outcome unlabeled 51,029. Removed: 0 SWE-bench rows, 0 duplicates, 0 with no assistant turn or unparsable. | |
| #### `trajectory-nemoswe_v2_oh` | |
| Source: [nvidia/Nemotron-SFT-SWE-v2](https://huggingface.co/datasets/nvidia/Nemotron-SFT-SWE-v2) at `bd151f3f2d` | |
| - The agentic subset of Nemotron-SFT-SWE-v2 (split `openhands_swe`, `data/swe.jsonl`): about 46k OpenHands trajectories generated by Qwen3-Coder-480B-A35B-Instruct and curated for SFT, on SWE-Gym and R2E-Gym-Subset issues. | |
| - The dataset's agentless subset (localization, repair and test generation by DeepSeek-R1-0528) is not in this config. | |
| - Label: No label field; all rows are `unlabeled`. | |
| - Used for: The card does not name a model trained on it (it lists papers, including Nemotron-Cascade, as references). | |
| - License: CC BY 4.0; Apache 2.0, MIT, BSD-3 and BSD-2 are listed as additional information. Ready for commercial use. | |
| - In this pool: 46,278 rows, 2.12B tokens (sigdig), 2.14B (SmolLM3); outcome unlabeled 46,278. Removed: 0 SWE-bench rows, 0 duplicates, 0 with no assistant turn or unparsable. | |
| #### `trajectory-scaleswe` | |
| Source: [AweAI-Team/Scale-SWE-Distilled](https://huggingface.co/datasets/AweAI-Team/Scale-SWE-Distilled) at `c14ce2130f` | |
| - Distillation trajectories from DeepSeek v3.2 on Scale-SWE task instances (the card: 71k trajectories, 3.5B tokens; the paper: 71,498 trajectories). Scale-SWE builds Python task instances from real GitHub pull requests. | |
| - The card does not name the harness, and does not say whether the trajectories are filtered to resolved runs. | |
| - Label: No label field; all rows are `unlabeled`. | |
| - Used for: The authors fine-tuned Qwen3-30B-A3B-Instruct on these trajectories (ScaleSWE Agent) and report 64% on SWE-bench Verified. | |
| - License: CC BY 4.0 (card text; the card has no license tag). | |
| - In this pool: 71,498 rows, 3.24B tokens (sigdig), 3.30B (SmolLM3); outcome unlabeled 71,498. Removed: 0 SWE-bench rows, 0 duplicates, 0 with no assistant turn or unparsable. | |
| #### `trajectory-sera` | |
| Source: [allenai/SERA-4.6-Lite-Best-Subset](https://huggingface.co/datasets/allenai/SERA-4.6-Lite-Best-Subset) at `ffcce5f9c2` | |
| - 47,498 trajectories from Sera-4.6-Lite-T1 and Sera-4.6-Lite-T2 (card). The authors kept only samples that the model submitted, then filtered by truncation ratio at 32K tokens (threshold 0.88). | |
| - The authors did not verify the trajectories; the source rows keep `rollout_patch` and `target_patch` (empty for T1) so that T2 rows can be checked. The card does not name the generator model or the harness. | |
| - Label: No label field; all rows are `unlabeled`. "Submitted" does not mean resolved. | |
| - Used for: The card reports that training on it reaches 50.67% +/- 1.86% on SWE-Bench Verified at 32K context; it does not name the model. | |
| - License: The card metadata says apache-2.0; the card text says ODC-By v1.0. Intended for research and educational use. | |
| - In this pool: 47,031 rows, 1.29B tokens (sigdig), 1.31B (SmolLM3); outcome unlabeled 47,031. Removed: 467 SWE-bench rows, 0 duplicates, 0 with no assistant turn or unparsable. | |
| ### Terminal and general agent tasks | |
| #### `trajectory-terminal` | |
| Source: [nvidia/Nemotron-Terminal-Corpus](https://huggingface.co/datasets/nvidia/Nemotron-Terminal-Corpus) at `67a3e039a1` | |
| - NVIDIA Terminal-Corpus: about 366k terminal-agent execution trajectories made with the Terminal-Task-Gen pipeline, in two streams: dataset adapters (about 226k; math, code and SWE datasets turned into terminal tasks) and skill-based synthetic tasks (about 140k; easy, medium and mixed). | |
| - The pool reads every parquet file of the repository; the `group` column keeps the folder. The card does not name the generator model or the harness. | |
| - Label: No label field; all rows are `unlabeled`. | |
| - Used for: The paper trains Nemotron-Terminal (from Qwen3 8B, 14B and 32B) on it and reports gains on Terminal-Bench 2.0; the card's table gives 27.4% +/- 2.4 for the 32B model against 3.4% +/- 1.6 for the Qwen3 base. The paper says most, not all, of its synthetic datasets are released. | |
| - License: CC BY 4.0 (card metadata). | |
| - In this pool: 366,154 rows, 6.08B tokens (sigdig), 6.07B (SmolLM3); outcome unlabeled 366,154. Removed: 0 SWE-bench rows, 0 duplicates, 0 with no assistant turn or unparsable. | |
| #### `trajectory-otagent` | |
| Source: [open-thoughts/OpenThoughts-Agent-SFT-100K](https://huggingface.co/datasets/open-thoughts/OpenThoughts-Agent-SFT-100K) at `45fb28fcc3` | |
| - The 100,000-row point of the OpenThoughts-Agent SFT scaling ladder: (task, agent trajectory) pairs. GLM-4.7-AWQ is the teacher in the terminus-2 harness inside Daytona sandboxes; traces are filtered to at least 5 model turns. | |
| - Tasks: SWE-Smith, StackExchange-SuperUser, StackExchange-Tezos (synthetically augmented) and IssueTasks. | |
| - Label: No label field; all rows are `unlabeled` (the source's `result` column is described only as rollout bookkeeping). | |
| - Used for: The card says it was used to fine-tune OpenThinkerAgent-8B-SFT-100K and OpenThinkerAgent-32B-SFT-100K. The 32B model card (from Qwen3-32B) reports 55.7 on SWE-Bench-Verified-100, 41.3 on OpenThoughts-TBLite and 26.2 on Terminal-Bench 2.0 with Terminus-2. | |
| - License: Apache 2.0. | |
| - In this pool: 94,334 rows, 1.46B tokens (sigdig), 1.46B (SmolLM3); outcome unlabeled 94,334. Removed: 0 SWE-bench rows, 0 duplicates, 0 with no assistant turn or unparsable. | |
| #### `trajectory-agenttrove` | |
| Source: [open-thoughts/AgentTrove](https://huggingface.co/datasets/open-thoughts/AgentTrove) at `b395a4307a` | |
| - Concatenation of 219 OpenThoughts-Agent training datasets (1,696,847 rows on the card): code repair, shell scripting, math, competitive programming and general computer-use tasks. All traces were generated with [Harbor](https://github.com/harbor-framework/harbor) and are released in the terminus-2 harness format. | |
| - Teacher models differ by source (examples on the card: GLM-4.7, GLM-4.6, GPT 5.1 Nano, Kimi K2.0 Thinking, MiniMax M2.0, Qwen3-8B). The `group` column keeps the source label (and the trace source when the row has one). | |
| - The card's source list has `-3pct` and `-10pct` variants, and one of its two `-reward1` variants, next to the same datasets without the suffix, so sources can overlap; the pool's exact deduplication removes repeated rows. | |
| - Rows with no assistant turn are removed (count below). | |
| - Label: The card documents a `reward` column (typically 1.0 success, 0.0 failure, missing for some sources). The files at this revision have a `result` column instead; the pool maps `1.0` to resolved, `0.0` to failed, error class names (for example `AgentTimeoutError`) to incomplete and empty values to unlabeled. This mapping is the pool's convention. | |
| - Used for: Released by the OpenThoughts-Agent team as a collection of agent traces; the card does not name a model trained on the whole collection. | |
| - License: Apache 2.0. | |
| - In this pool: 1,330,717 rows, 14.80B tokens (sigdig), 14.77B (SmolLM3); outcome unlabeled 1,132,063, incomplete 189,440, resolved 8,058, failed 1,156. Removed: 0 SWE-bench rows, 237,695 duplicates, 128,435 with no assistant turn or unparsable. | |
| #### `trajectory-opencode` | |
| Source: [nvidia/Nemotron-SFT-OpenCode-v1](https://huggingface.co/datasets/nvidia/Nemotron-SFT-OpenCode-v1) at `556d5237ac` | |
| - NVIDIA agentic instruction-tuning data for the [OpenCode](https://opencode.ai/) CLI: about 459K traces in six subsets (general, bash_only_tool, bash_only_tool_skills, question_tool, agent_skills, agent_skills_question_tool); the `group` column keeps the subset. | |
| - Questions and agent skills are synthetic (GPT OSS 120B); solution traces are from Qwen3-Coder-480B-A35B-Instruct. In the question_tool subset, the user's replies are generated by the same model as the rollout (the card does not say how agent_skills_question_tool replies are made). | |
| - Label: No label field; all rows are `unlabeled`. | |
| - Used for: The card does not name a model trained on it. | |
| - License: CC BY 4.0; ready for commercial and non-commercial use. | |
| - In this pool: 460,251 rows, 5.87B tokens (sigdig), 5.87B (SmolLM3); outcome unlabeled 460,251. Removed: 0 SWE-bench rows, 3 duplicates, 0 with no assistant turn or unparsable. | |
| ## Licenses | |
| Each config keeps the license of its source dataset. The trajectories quote code from public repositories, which | |
| stays under the licenses of those repositories. Check the source cards before use. | |
| | Config | Source | License as stated by the source | | |
| |---|---|---| | |
| | `trajectory-nanoswe` | [nanoswe/nanoswe-trajs-261002](https://huggingface.co/datasets/nanoswe/nanoswe-trajs-261002) | MIT. Code quoted in tool observations stays under the licenses of the upstream projects. | | |
| | `trajectory-minicoder` | [ricdomolm/mini-coder-trajs-400k](https://huggingface.co/datasets/ricdomolm/mini-coder-trajs-400k) | MIT (card metadata). | | |
| | `trajectory-klear` | [Kwai-Klear/SWE-smith-mini_swe_agent_plus-trajectories-66k](https://huggingface.co/datasets/Kwai-Klear/SWE-smith-mini_swe_agent_plus-trajectories-66k) | MIT (card metadata). | | |
| | `trajectory-openswe` | [nvidia/Open-SWE-Traces](https://huggingface.co/datasets/nvidia/Open-SWE-Traces) | CC BY 4.0; ready for commercial and non-commercial use. Each row of the source records the SPDX license of its repository. | | |
| | `trajectory-coderforge` | [togethercomputer/CoderForge-Preview](https://huggingface.co/datasets/togethercomputer/CoderForge-Preview) | No dataset license is stated on the card. Rows of the source carry a `license` field that the card does not describe. | | |
| | `trajectory-nebius_oh` | [nebius/SWE-rebench-openhands-trajectories](https://huggingface.co/datasets/nebius/SWE-rebench-openhands-trajectories) | CC BY 4.0. | | |
| | `trajectory-swezero` | [nvidia/SWE-Zero-openhands-trajectories](https://huggingface.co/datasets/nvidia/SWE-Zero-openhands-trajectories) | CC BY 4.0; ready for commercial and non-commercial use. The card states that only repositories under MIT, Apache-2.0, BSD-2-Clause and BSD-3-Clause are used. | | |
| | `trajectory-nemoswe_v1` | [nvidia/Nemotron-SWE-v1](https://huggingface.co/datasets/nvidia/Nemotron-SWE-v1) | CC BY 4.0, except certain subsets under BSD-3-Clause as the Hugging Face Dataset Viewer shows; Apache 2.0 and MIT are listed as additional information. Ready for commercial use. | | |
| | `trajectory-nemoswe_v2_oh` | [nvidia/Nemotron-SFT-SWE-v2](https://huggingface.co/datasets/nvidia/Nemotron-SFT-SWE-v2) | CC BY 4.0; Apache 2.0, MIT, BSD-3 and BSD-2 are listed as additional information. Ready for commercial use. | | |
| | `trajectory-scaleswe` | [AweAI-Team/Scale-SWE-Distilled](https://huggingface.co/datasets/AweAI-Team/Scale-SWE-Distilled) | CC BY 4.0 (card text; the card has no license tag). | | |
| | `trajectory-sera` | [allenai/SERA-4.6-Lite-Best-Subset](https://huggingface.co/datasets/allenai/SERA-4.6-Lite-Best-Subset) | The card metadata says apache-2.0; the card text says ODC-By v1.0. Intended for research and educational use. | | |
| | `trajectory-terminal` | [nvidia/Nemotron-Terminal-Corpus](https://huggingface.co/datasets/nvidia/Nemotron-Terminal-Corpus) | CC BY 4.0 (card metadata). | | |
| | `trajectory-otagent` | [open-thoughts/OpenThoughts-Agent-SFT-100K](https://huggingface.co/datasets/open-thoughts/OpenThoughts-Agent-SFT-100K) | Apache 2.0. | | |
| | `trajectory-agenttrove` | [open-thoughts/AgentTrove](https://huggingface.co/datasets/open-thoughts/AgentTrove) | Apache 2.0. | | |
| | `trajectory-opencode` | [nvidia/Nemotron-SFT-OpenCode-v1](https://huggingface.co/datasets/nvidia/Nemotron-SFT-OpenCode-v1) | CC BY 4.0; ready for commercial and non-commercial use. | | |
| ## Build | |
| `scripts/build_pool.py` in this repository builds each config from the source revision in its section. The | |
| per-config numbers are in `stats/trajectory/<source>.json`. | |
| ## References | |
| Papers and posts cited by the source cards: | |
| - CoderForge-Preview: SOTA Open Dataset for Training Efficient Agents (Together AI blog, February 2026, https://www.together.ai/blog/coderforge-preview) | |
| - Computational Arbitrage in AI Model Markets (cited by the card; arXiv 2603.22404 per the mini-coder model cards) | |
| - From SWE-ZERO to SWE-HERO: Execution-free to Execution-based Fine-tuning for Software Engineering Agents (arXiv 2604.01496) | |
| - Immersion in the GitHub Universe: Scaling Coding Agents to Mastery (arXiv 2602.09892) | |
| - OPEN-SWE-TRACES: Advancing Dual-Mode Multilingual Distillation for Software Engineering Agents (Findings of EMNLP 2026, https://aclanthology.org/2026.findings-emnlp.131/) | |
| - On Data Engineering for Scaling LLM Terminal Capabilities (arXiv 2602.21193) | |
| - OpenHands Trajectories with Qwen3-Coder-480B-A35B-Instruct (Nebius blog, 2025) | |
| - OpenThoughts-Agent (OpenThoughts blog, December 2025) | |
| - OpenThoughts-Agent: Data Recipes for Agentic Models (OpenThoughts blog, 2026) | |