datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
nemotron-post-training-v2-qwen-3.5-9b-regen
Dataset Card for Nemotron Post Training v2 Qwen 3.5 9B Regen
Regenerated responses from nvidia/Nemotron-Post-Training-Dataset-v2 dataset using Qwen3.5 9B model.
Parameter
Value
Max Tokens
4096
Temperature
1.0
Top-k
20
Top-p
0.95
Repetition Penalty
1.5
Dataset consists only the english samples from the Nemotron Post Training Dataset. 85% of the chat prompts have reasoning enabled, every other category has reasoning disabled.
Category
Value
math… See the full description on the dataset page: https://huggingface.co/datasets/Dogacel/nemotron-post-training-v2-qwen-3.5-9b-regen.qwen-9b-3m
qwen-9b-3m
Multi-domain SFT-target dataset: ~3,000,000 prompts, each with ONE completion
generated offline by Qwen/Qwen3.5-9B (thinking mode). Exported snapshot from an
offline queue pipeline; repartitioned into 512 parquet shards.
Columns (21)
record_index, input_sha256, prompt_sha256, dataset, split, source, upstream_id,
bucket, messages_json, prompt, prompt_token_count, generation_seed, enable_thinking,
worker, executor_worker, completion, completion_input_ids… See the full description on the dataset page: https://huggingface.co/datasets/dipta007/qwen-9b-3m.harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-30m-historical-20t-think
harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-30m-historical-20t-think
1,000 historical evaluation attempts (250 tasks, four samples per task), newly
graded with gpt-5.6-sol using Harvey's original per-criterion rubric prompt
and all-criteria-pass rule. Mean all-pass rate: 5.0000%.
The train split contains evaluation records, not training examples.
Generation and grading protocols
Generation is unchanged: historical 20-turn thinking-enabled
glob/grep/read agent… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-30m-historical-20t-think.harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-3m-historical-20t-think
harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-3m-historical-20t-think
1,000 historical evaluation attempts (250 tasks, four samples per task), newly
graded with gpt-5.6-sol using Harvey's original per-criterion rubric prompt
and all-criteria-pass rule. Mean all-pass rate: 1.3000%.
The train split contains evaluation records, not training examples.
Generation and grading protocols
Generation is unchanged: historical 20-turn thinking-enabled
glob/grep/read agent… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-3m-historical-20t-think.harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-10m-historical-20t-think
harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-10m-historical-20t-think
1,000 historical evaluation attempts (250 tasks, four samples per task), newly
graded with gpt-5.6-sol using Harvey's original per-criterion rubric prompt
and all-criteria-pass rule. Mean all-pass rate: 4.0000%.
The train split contains evaluation records, not training examples.
Generation and grading protocols
Generation is unchanged: historical 20-turn thinking-enabled
glob/grep/read agent… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-10m-historical-20t-think.qwen9b-coop-claude-code
qwen9b-coop-claude-code
Two-agent cooperative coding trajectories generated by running
CooperBench in coop mode on
the CooperData task set, using
Qwen/Qwen3.5-9B as the model and Claude Code (claude_code) as the
agent framework. Each pair runs two agents in parallel — one per feature —
coordinating via Redis messaging and a shared git remote.
The matched solo (single-agent) baseline is at
CooperBench/qwen9b-solo-claude-code.
Same task corpus, same model, same agent — only the… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/qwen9b-coop-claude-code.qwen3.5-9B-tau2bench-retail-traces
Qwen3.5-9B τ²-bench retail traces (judged)
Curated retail-domain traces collected on tau2-bench by running Qwen3.5-9B
(with and without memory-rule injection) on the canonical 114-task retail
pool. Each trace is judged by two independent signals:
Canonical tau2-bench evaluator (canonical_reward ∈ {{0, 1}})
— the official task-success metric, combining a DB-state check and
gpt-4.1-2025-04-14 NL-assertion verifier.
Blind process-quality judge (judge_retail.*)
— a retail-shaped… See the full description on the dataset page: https://huggingface.co/datasets/KermitCO/qwen3.5-9B-tau2bench-retail-traces.SWEbench-Verified-eval150-M2.7-Qwen3.5-9B-orch-7arms-2repeats-w32-20260920
SWE-bench Verified eval150 — M2.7 × Qwen3.5-9B, seven arms, two repeats, 32 concurrency
Campaign 2026-09-20. 14/14 independent full150 runs audited. Complete accuracy evidence.
Evaluation mode is orch: MiniMax-M2.7 orchestrator and the specified Qwen3.5-9B worker. Training mode is labeled independently. All runs use 32 concurrent episodes, 10GiB Docker sandboxes, four TP1 workers and one TP4/EP4 coordinator. Frozen regression-gated prompts, decoding and canonical verifier match… See the full description on the dataset page: https://huggingface.co/datasets/CharlieLLL/SWEbench-Verified-eval150-M2.7-Qwen3.5-9B-orch-7arms-2repeats-w32-20260920.harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-1m-historical-20t-think
harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-1m-historical-20t-think
1,000 historical evaluation attempts (250 tasks, four samples per task), newly
graded with gpt-5.6-sol using Harvey's original per-criterion rubric prompt
and all-criteria-pass rule. Mean all-pass rate: 2.2000%.
The train split contains evaluation records, not training examples.
Generation and grading protocols
Generation is unchanged: historical 20-turn thinking-enabled
glob/grep/read agent… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-gpt56sol-qwen35-9b-notes70-recall30-1m-historical-20t-think.tmax-tasks-selfgen-qwen35-9b-20260919-1k-verified
tmax self-generated tasks — Qwen3.5-9B (verified arm)
The 1,042 tasks from
…-20260919-1k,
each graded against the issue-#12 rubric by the same model that generated them
(hamishivi/Qwen3.5-9B). Using the generator as its own reviewer is deliberate: the
question is whether an open-weights model can carry both halves of the loop. A stronger
reviewer would answer a different question.
The grader sees instruction / setup.sh / tests only. truth is withheld from it,
so it is no better… See the full description on the dataset page: https://huggingface.co/datasets/osieosie/tmax-tasks-selfgen-qwen35-9b-20260919-1k-verified.Angry-Claudius-9B-Dataset
Angry Claudius 9B Dataset
The training and evaluation data used to develop
Angry Claudius 9B, a
joke model trained to answer user requests with short profane dismissals instead
of completing the requested task.
Content warning
This dataset contains frequent explicit profanity. It is intended for behavioral
fine-tuning and evaluation research and is unsuitable for applications that require
polite, helpful, or family-friendly responses.
Data… See the full description on the dataset page: https://huggingface.co/datasets/axiomofmind/Angry-Claudius-9B-Dataset.harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-kl-0p05-think
harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-kl-0p05-think
Complete closed-book recall evaluation: 7,933 probes. One dataset repository for Qwen3.5-9B, the 100M notes + note-conditioned trajectory mixture, and KL coefficient 0.05.
The train split contains evaluation records. Each row is one scored probe; this split name follows the existing evaluation dataset layout.
Model, data, and KL condition
Evaluated model:… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-kl-0p05-think.harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-kl-0p01-think
harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-kl-0p01-think
Complete closed-book recall evaluation: 7,933 probes. One dataset repository for Qwen3.5-9B, the 100M notes + note-conditioned trajectory mixture, and KL coefficient 0.01.
The train split contains evaluation records. Each row is one scored probe; this split name follows the existing evaluation dataset layout.
Model, data, and KL condition
Evaluated model:… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-kl-0p01-think.harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-kl-0p1-think
harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-kl-0p1-think
Complete closed-book recall evaluation: 7,933 probes. One dataset repository for Qwen3.5-9B, the 100M notes + note-conditioned trajectory mixture, and KL coefficient 0.1.
The train split contains evaluation records. Each row is one scored probe; this split name follows the existing evaluation dataset layout.
Model, data, and KL condition
Evaluated model:… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-kl-0p1-think.tmax-tasks-selfgen-qwen35-9b-20260919-1k
tmax self-generated tasks — Qwen3.5-9B (raw arm)
1,042 agentic terminal tasks generated by hamishivi/Qwen3.5-9B using the rl_data
pipeline. Every previous tmax corpus was written by an API model (gemini-3.1-pro-preview);
this one asks whether the model we RL on can generate its own training data.
Companion: …-1k-verified
— the same 1,042 tasks with rubric labels from the same model acting as reviewer.
Recipe
Identical to the Gemini 1k run… See the full description on the dataset page: https://huggingface.co/datasets/osieosie/tmax-tasks-selfgen-qwen35-9b-20260919-1k.harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-10m-think
harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-10m-think
Complete closed-book recall evaluation: 7,933 probes. One dataset repository for Qwen3.5-9B, the 10M notes + note-conditioned trajectory mixture, and no KL regularization.
The train split contains evaluation records. Each row is one scored probe; this split name follows the existing evaluation dataset layout.
Model, data, and KL condition
Evaluated model:… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-10m-think.harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-30m-think
harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-30m-think
Complete closed-book recall evaluation: 7,933 probes. One dataset repository for Qwen3.5-9B, the 30M notes + note-conditioned trajectory mixture, and no KL regularization.
The train split contains evaluation records. Each row is one scored probe; this split name follows the existing evaluation dataset layout.
Model, data, and KL condition
Evaluated model:… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-30m-think.qwen9b-solo-claude-code
qwen9b-solo-claude-code
Single-agent coding trajectories generated by running
CooperBench in solo mode on
the CooperData task set, using
Qwen/Qwen3.5-9B as the model and Claude Code (claude_code) as the
agent framework. One agent implements both features in each task.
The matched coop (two-agent) version is at
CooperBench/qwen9b-coop-claude-code.
Same task corpus, same model, same agent — only the coordination differs, so
together they isolate the cooperation deficit.
At a… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/qwen9b-solo-claude-code.harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-think
harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-think
Complete closed-book recall evaluation: 7,933 probes. One dataset repository for Qwen3.5-9B, the 100M notes + note-conditioned trajectory mixture, and no KL regularization.
The train split contains evaluation records. Each row is one scored probe; this split name follows the existing evaluation dataset layout.
Model, data, and KL condition
Evaluated model:… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-100m-think.harvey-eval-gpt56sol-qwen35-9b-base-20t-think
harvey-eval-gpt56sol-qwen35-9b-base-20t-think
1,000 historical evaluation attempts (250 tasks, four samples per task), newly
graded with gpt-5.6-sol using Harvey's original per-criterion rubric prompt
and all-criteria-pass rule. Mean all-pass rate: 3.9000%.
The train split contains evaluation records, not training examples.
Generation and grading protocols
Generation is unchanged: historical 20-turn thinking-enabled
glob/grep/read agent runs. These are not new… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-gpt56sol-qwen35-9b-base-20t-think.harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-1m-think
harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-1m-think
Complete closed-book recall evaluation: 7,933 probes. One dataset repository for Qwen3.5-9B, the 1M notes + note-conditioned trajectory mixture, and no KL regularization.
The train split contains evaluation records. Each row is one scored probe; this split name follows the existing evaluation dataset layout.
Model, data, and KL condition
Evaluated model:… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-recall-qwen35-9b-notes70-notecondtraj30-1m-think.harvey-eval-gpt56sol-qwen35-9b-notes70-notecondtraj30-1m-historical-20t-think
harvey-eval-gpt56sol-qwen35-9b-notes70-notecondtraj30-1m-historical-20t-think
1,000 historical evaluation attempts (250 tasks, four samples per task), graded
with gpt-5.6-sol using Harvey's original per-criterion rubric prompt and
binary all-criteria-pass rule. Mean all-pass rate: 3.1000%.
The train split contains held-out evaluation records, not training examples.
Model and training mixture
The evaluated checkpoint is Qwen3.5-9B trained for two epochs on the… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-gpt56sol-qwen35-9b-notes70-notecondtraj30-1m-historical-20t-think.harvey-eval-gpt56sol-qwen35-9b-notes70-notecondtraj30-5m-historical-20t-think
harvey-eval-gpt56sol-qwen35-9b-notes70-notecondtraj30-5m-historical-20t-think
1,000 historical evaluation attempts (250 tasks, four samples per task), graded
with gpt-5.6-sol using Harvey's original per-criterion rubric prompt and
binary all-criteria-pass rule. Mean all-pass rate: 3.4000%.
The train split contains held-out evaluation records, not training examples.
Model and training mixture
The evaluated checkpoint is Qwen3.5-9B trained for two epochs on the… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-gpt56sol-qwen35-9b-notes70-notecondtraj30-5m-historical-20t-think.harvey-eval-gpt56sol-qwen35-9b-notes70-notecondtraj30-30m-historical-20t-think
harvey-eval-gpt56sol-qwen35-9b-notes70-notecondtraj30-30m-historical-20t-think
1,000 historical evaluation attempts (250 tasks, four samples per task), graded
with gpt-5.6-sol using Harvey's original per-criterion rubric prompt and
binary all-criteria-pass rule. Mean all-pass rate: 7.0000%.
The train split contains held-out evaluation records, not training examples.
Model and training mixture
The evaluated checkpoint is Qwen3.5-9B trained for two epochs on the… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-gpt56sol-qwen35-9b-notes70-notecondtraj30-30m-historical-20t-think.qwen9b-coop-mini-swe-agent
qwen9b-coop-mini-swe-agent
Two-agent cooperative coding trajectories generated by running
CooperBench in coop mode on
the CooperData task set, using
Qwen/Qwen3.5-9B as the model and mini_swe_agent_v2 as the agent framework.
Each pair runs two agents in parallel — one per feature — coordinating via Redis messaging and a shared git remote.
The matched solo version is at
CooperBench/qwen9b-solo-mini-swe-agent.
Same task corpus, same model, same agent — only the coordination differs… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/qwen9b-coop-mini-swe-agent.harvey-eval-gpt56sol-qwen35-9b-notes70-notecondtraj30-100m-historical-20t-think
harvey-eval-gpt56sol-qwen35-9b-notes70-notecondtraj30-100m-historical-20t-think
1,000 historical evaluation attempts (250 tasks, four samples per task), graded
with gpt-5.6-sol using Harvey's original per-criterion rubric prompt and
binary all-criteria-pass rule. Mean all-pass rate: 8.0000%.
The train split contains held-out evaluation records, not training examples.
Model and training mixture
The evaluated checkpoint is Qwen3.5-9B trained for two epochs on the… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-gpt56sol-qwen35-9b-notes70-notecondtraj30-100m-historical-20t-think.qwen9b-solo-mini-swe-agent
qwen9b-solo-mini-swe-agent
Single-agent coding trajectories generated by running
CooperBench in solo mode on
the CooperData task set, using
Qwen/Qwen3.5-9B as the model and mini_swe_agent_v2 as the agent framework.
One agent implements both features in each task.
The matched coop version is at
CooperBench/qwen9b-coop-mini-swe-agent.
Same task corpus, same model, same agent — only the coordination differs, so
together they isolate the cooperation deficit.
At a glance… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/qwen9b-solo-mini-swe-agent.harvey-eval-gpt56sol-qwen35-9b-notes70-notecondtraj30-10m-historical-20t-think
harvey-eval-gpt56sol-qwen35-9b-notes70-notecondtraj30-10m-historical-20t-think
1,000 historical evaluation attempts (250 tasks, four samples per task), graded
with gpt-5.6-sol using Harvey's original per-criterion rubric prompt and
binary all-criteria-pass rule. Mean all-pass rate: 6.2000%.
The train split contains held-out evaluation records, not training examples.
Model and training mixture
The evaluated checkpoint is Qwen3.5-9B trained for two epochs on the… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/harvey-eval-gpt56sol-qwen35-9b-notes70-notecondtraj30-10m-historical-20t-think.ch-pilot-rollouts-qwen3.5-9b
C&H Pilot Rollouts — Qwen/Qwen3.5-9B
20 agentic exploration rollouts over the Calderwood & Harkness (C&H) synthetic law-firm
corpus (the open-sourced world from harvey-labs
tasks/firm-knowledge/, MIT), generated by Qwen/Qwen3.5-9B served with vLLM.
Part of an actor-selection pilot for a world-internalization research project: the goal is to
mine agent trajectories into verified fact stores and rewritten likelihood-training targets.
Companion dataset (same seeds/tasks, different… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/ch-pilot-rollouts-qwen3.5-9b.wmrl-v4-base9b-agentic-eval-20t-think
Base-model agentic eval transcripts, 20-turn budget (WM-RL v4)
1,000 complete agentic-evaluation transcripts of the untrained base model Qwen/Qwen3.5-9B (revision
c202236235762e1c871ad0ccb60c8ee5ba337b9a), thinking enabled, on the 250 held-out firm-knowledge tasks of the WM-RL v4
study, at a 20-turn tool budget. This is the baseline every trained condition in the study is compared
against; the transcripts are the raw rollouts, saved before grading.
250 tasks x 4 samples = 1,000… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/wmrl-v4-base9b-agentic-eval-20t-think.