conversations listlengths 2 138 | n_tokens int64 78 32.8k | agent stringclasses 0
values | completed stringclasses 0
values | dataset stringclasses 1
value | episode stringclasses 0
values | expected_answer stringclasses 768
values | hf_repo stringclasses 3
values | hf_revision stringclasses 1
value | injected_classes stringclasses 375
values | judgment stringclasses 0
values | kind stringclasses 2
values | license stringclasses 2
values | loss_mask_note stringclasses 1
value | meta_json stringlengths 294 329 ⌀ | metadata stringlengths 2 427 ⌀ | mix_source stringclasses 4
values | model stringclasses 0
values | n_chars stringlengths 3 6 ⌀ | n_messages stringclasses 21
values | n_turns stringclasses 39
values | not_errored_cut stringclasses 0
values | original_source stringclasses 11
values | original_teacher stringclasses 4
values | problem stringclasses 780
values | recovery_ish stringclasses 1
value | result stringclasses 0
values | run_id stringclasses 0
values | shard stringclasses 6
values | source stringclasses 1
value | subset stringclasses 6
values | task_complete_in_last stringclasses 0
values | trial_name stringlengths 18 41 ⌀ | upstream_columns stringclasses 3
values | upstream_source stringclasses 8
values | uuid stringlengths 32 42 ⌀ | verified stringclasses 0
values |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
[
{
"role": "user",
"content": "You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be given a task description and the output from previously executed commands. Your goal is to solve the task by providing batches of shell commands.\n\nFormat your response as JSON w... | 27,368 | null | null | null | null | null | null | null | [{"at": 9, "classes": ["prose", "prose", "prose"], "malformed_train_loss": 0, "recovery_train_loss": 1}, {"at": 17, "classes": ["role_leak", "role_leak"], "malformed_train_loss": 0, "recovery_train_loss": 1}] | null | null | null | train_loss=0 on injected failures and all user turns; 1 on recovery and other teacher assistant turns (upstream train_loss kept); consumed by SkyRL tokenize | null | null | recovery | null | 100030 | null | 23 | null | synthetic_structural_recovery | Kimi-2.5 | null | true | null | null | null | null | null | null | r2egym-2457__PSm4dhy | null | r2egym | null | null |
[{"role":"user","content":"You are an AI assistant tasked with solving command-line tasks in a Linux(...TRUNCATED) | 20,900 | null | null | null | null | null | null | null | "[{\"at\": 3, \"classes\": [\"role_leak\", \"role_leak\", \"role_leak\"], \"malformed_train_loss\": (...TRUNCATED) | null | null | null | "train_loss=0 on injected failures and all user turns; 1 on recovery and other teacher assistant tur(...TRUNCATED) | null | null | recovery | null | 61187 | null | 25 | null | synthetic_structural_recovery | GLM-4.7 | null | true | null | null | null | null | null | null | swesmith-06915__pUGnqmx | null | swesmith | null | null |
[{"role":"user","content":"You are an AI assistant tasked with solving command-line tasks in a Linux(...TRUNCATED) | 7,990 | null | null | null | null | null | null | null | "[{\"at\": 7, \"classes\": [\"commands_only\", \"commands_only\"], \"malformed_train_loss\": 0, \"re(...TRUNCATED) | null | null | null | "train_loss=0 on injected failures and all user turns; 1 on recovery and other teacher assistant tur(...TRUNCATED) | null | null | recovery | null | 23920 | null | 13 | null | synthetic_structural_recovery | GLM-4.7 | null | true | null | null | null | null | null | null | staqc_embedding_mean_filtered-48__qbBFrZj | null | staqc | null | null |
[{"role":"user","content":"You are an AI assistant tasked with solving command-line tasks in a Linux(...TRUNCATED) | 8,389 | null | null | null | null | null | null | null | "[{\"at\": 15, \"classes\": [\"tool_call_content\", \"tool_call_content\", \"tool_call_content\"], \(...TRUNCATED) | null | null | null | "train_loss=0 on injected failures and all user turns; 1 on recovery and other teacher assistant tur(...TRUNCATED) | null | null | recovery | null | 27623 | null | 18 | null | synthetic_structural_recovery | Kimi-2.5 | null | true | null | null | null | null | null | null | swesmith-05287__7UnHVrv | null | swesmith | null | null |
[{"role":"user","content":"You are an AI assistant tasked with solving command-line tasks in a Linux(...TRUNCATED) | 21,429 | null | null | null | null | null | null | null | "[{\"at\": 9, \"classes\": [\"prose\", \"prose\", \"prose\"], \"malformed_train_loss\": 0, \"recover(...TRUNCATED) | null | null | null | "train_loss=0 on injected failures and all user turns; 1 on recovery and other teacher assistant tur(...TRUNCATED) | null | null | recovery | null | 71344 | null | 17 | null | synthetic_structural_recovery | MiniMax M2.0 | null | true | null | null | null | null | null | null | overflow-4591__7vg89be | null | StackExchange Overflow | null | null |
[{"role":"user","content":"You are an AI assistant tasked with solving command-line tasks in a Linux(...TRUNCATED) | 8,469 | null | null | null | null | null | null | null | "[{\"at\": 7, \"classes\": [\"tool_call_content\", \"tool_call_content\"], \"malformed_train_loss\":(...TRUNCATED) | null | null | null | "train_loss=0 on injected failures and all user turns; 1 on recovery and other teacher assistant tur(...TRUNCATED) | null | null | recovery | null | 28682 | null | 11 | null | synthetic_structural_recovery | MiniMax M2.0 | null | true | null | null | null | null | null | null | overflow-3805__Tn2EZRt | null | StackExchange Overflow | null | null |
[{"role":"user","content":"You are an AI assistant tasked with solving command-line tasks in a Linux(...TRUNCATED) | 25,386 | null | null | null | null | null | null | null | "[{\"at\": 9, \"classes\": [\"role_leak\", \"commands_only\"], \"malformed_train_loss\": 0, \"recove(...TRUNCATED) | null | null | null | "train_loss=0 on injected failures and all user turns; 1 on recovery and other teacher assistant tur(...TRUNCATED) | null | null | recovery | null | 74743 | null | 20 | null | synthetic_structural_recovery | GLM-4.7 | null | true | null | null | null | null | null | null | issue-4299__AsP7JdZ | null | swesmith | null | null |
[{"role":"user","content":"You are an AI assistant tasked with solving command-line tasks in a Linux(...TRUNCATED) | 3,905 | null | null | null | null | null | null | null | "[{\"at\": 5, \"classes\": [\"tool_call_content\", \"tool_call_content\", \"tool_call_content\"], \"(...TRUNCATED) | null | null | null | "train_loss=0 on injected failures and all user turns; 1 on recovery and other teacher assistant tur(...TRUNCATED) | null | null | recovery | null | 11077 | null | 11 | null | synthetic_structural_recovery | GLM-4.7 | null | true | null | null | null | null | null | null | dockerfile-2185__jgwrVq4 | null | taskmaster | null | null |
[{"role":"user","content":"You are an AI assistant tasked with solving command-line tasks in a Linux(...TRUNCATED) | 9,421 | null | null | null | null | null | null | null | "[{\"at\": 3, \"classes\": [\"role_leak\"], \"malformed_train_loss\": 0, \"recovery_train_loss\": 1}(...TRUNCATED) | null | null | null | "train_loss=0 on injected failures and all user turns; 1 on recovery and other teacher assistant tur(...TRUNCATED) | null | null | recovery | null | 28699 | null | 10 | null | synthetic_structural_recovery | GLM-4.7 | null | true | null | null | null | null | null | null | staqc_embedding_mean_filtered-73__f9fsKhR | null | staqc | null | null |
[{"role":"user","content":"You are an AI assistant tasked with solving command-line tasks in a Linux(...TRUNCATED) | 21,395 | null | null | null | null | null | null | null | "[{\"at\": 23, \"classes\": [\"role_leak\"], \"malformed_train_loss\": 0, \"recovery_train_loss\": 1(...TRUNCATED) | null | null | null | "train_loss=0 on injected failures and all user turns; 1 on recovery and other teacher assistant tur(...TRUNCATED) | null | null | recovery | null | 57205 | null | 22 | null | synthetic_structural_recovery | GLM-4.7 | null | true | null | null | null | null | null | null | tezos-0159_copy1__UbusoKG | null | stack_exchange | null | null |
Grug SFT mix v4, thinking (40M tokens)
A 40M-token SFT mix for Grug-67B-A2B (open-athena Datakit SFT 2026-09-21) in which the teachers' reasoning is kept in Grug's think channel. It fixes Grug's terminus-2 agent failures (malformed actions, no recovery after a parse error, weak context-summarization handoffs) without removing its reasoning.
7,453 conversations, 39.9M tokens (Grug tokenizer, its chat template with Reasoning: /think), of which 12.9M are
trained. 99.7% of trained agent turns carry a think block.
| component | rows | trained turns | with think block | trained tokens |
|---|---|---|---|---|
agentic — AgentTrove clean completed trajectories |
2,186 | 19,347 | 99.7% | 8.33M |
recovery — synthetic structural recovery |
400 | 3,950 | 99.7% | 1.72M |
handoff — AgentTrove context-summarization handoffs |
441 | 1,028 | 99.6% | 0.81M |
non_agentic — Nemotron IF / math / science |
4,426 | 5,785 | 0% | 2.05M |
Result
SFT of Grug 09.21 on this mix (SkyRL main_sft, Megatron, lr 1e-5 cosine, batch 32; step 300 ≈ 1.3 epochs),
evaluated with terminus-2 at 3× task timeouts:
| benchmark | base, thinking | base, no thinking | SFT step 300 |
|---|---|---|---|
| SWE-bench Verified, random 100 | 19.3 (mean of 4) | 15 | 40 |
| Terminal-Bench 2.1, 74 validated tasks | 5 | 5 | 9 |
| TBLite 100 (any reward) | 10 | 20 | 24 |
| terminus format errors (share of turns, SWE-bench) | 82.6% | 51.9% | 1.5% |
The trained model always opens a think block, also under /nothink.
Format
data/train-00000-of-00001.parquet, one conversation per row:
conversations:list<struct<role, content, reasoning_content, train_loss: int8>>.reasoning_contentis the teacher's reasoning for an assistant turn; Grug's chat template renders it as<|start_think|>…<|end_think|>beforecontent, which holds the terminus-2 JSON action alone.train_lossis 1 on assistant turns to train on and 0 on everything else: user / tool turns, assistant turns the terminus-2 harness rejected, injected malformed turns in recovery rows, and teacher turns that still carried text around the action or tool-call / role markup.
mix_source:recovery | handoff | agentic | non_agentic;n_tokens: templated length before the final cleanup.- Provenance (strings, may be null):
original_source,original_teacher,trial_name,kind,upstream_source,hf_repo,license,injected_classes,meta_json, ...
Every row ends with an assistant turn and is at most 32,768 tokens.
Train with loss on all assistant turns while honouring train_loss, rendering with the model's own template
(enable_thinking true or unset), e.g. SkyRL main_sft (marianna13/SkyRL@301659a8 or later):
messages_key=conversations train_on_what=all_assistant_messages.
Components
agentic — open-thoughts/AgentTrove@b395a430
terminus-2 trajectories, clean-completion cut (exp_* / unknown sources dropped, last turn a complete non-error
task_complete, longest episode per trial and run, ≤20k rows per source), restricted to teachers that reason in a
think block: GLM-4.7, GLM-5.0, Kimi-2.5, Kimi K2.0 Thinking, MiniMax M2.0. A teacher's <think>…</think> block (or
GLM-4.7's form with only the closing tag) becomes reasoning_content. Assistant turns are trained only if the
original text parses as terminus-2 JSON.
handoff — AgentTrove GLM-4.7 rows around a context-summarization handoff (≤3 trials per task): summary
(only the summary is trained), answers (only the answers are trained), continuation (the new agent's questions and
turns after the handoff).
recovery — clean agentic trajectories with 1–2 injection points each. Before a teacher turn, 1–3 failed attempts
are inserted, each a re-rendering of that teacher turn in one of Grug's measured error styles, followed by the
harness's real parse-error message; the teacher's valid turn (with its reasoning) follows as the trained recovery.
No Grug output text is used. Trajectories used here are excluded from agentic.
non_agentic — Nemotron post-training v3 SFT sets (nvidia/Nemotron-SFT-Math-v4, nvidia/Nemotron-SFT-Science-v2,
nvidia/Nemotron-SFT-Instruction-Following-Chat-v3), reasoning stripped, as an anti-forgetting slice. These rows
have no think block.
Build
In marianna13/monorepo (issue 0048): rl/scripts/sft_gen/agenttrove_convert.py --think-to-reasoning, then rl/scripts/data/build_grug_think_mix.sh (recovery synthesis, token-balanced sampling with
build_sft_mix_5k.py --balance tokens --total-tokens 40000000 --enable-thinking true --fractions recovery:0.12,handoff:0.25,agentic:0.55,non_agentic:0.08, seed 50, then thinkify_mix.py). mix_manifest.json
records the sampling (7,460 rows; thinkify_mix.py then dropped 7 rows left with no trained turn).
Limitations
- The
non_agenticrows carry no reasoning, so for chat and math questions they teach/thinkfollowed by no think block. - AgentTrove teacher traces may overlap other public SFT mixes. No row mentions any instance of the SWE-bench Verified random-100 subset used above, or contains a slice of its problem statements; this was not checked for other benchmarks.
- The teacher set (GLM, Kimi, MiniMax) differs from the v2 mix (mostly GLM-4.6), so v2 and v4 differ in more than reasoning.
License
Released under CC BY-SA 4.0 because it contains CC BY-SA 4.0 Nemotron rows (Science-v2, Math-v4); the AgentTrove-derived rows are Apache-2.0 upstream. Teacher-model outputs remain subject to their providers' terms.
- Downloads last month
- 50