Datasets:
File size: 3,508 Bytes
a5d2163 c0ec3ad a5d2163 c0ec3ad | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 | ---
pretty_name: Qwen3-4B Teacher Rollouts 76K Non-Thinking
language:
- en
task_categories:
- text-generation
tags:
- reasoning
- knowledge-distillation
- teacher-rollout
- qwen3
size_categories:
- 10K<n<100K
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
---
# Qwen3-4B Teacher Rollouts 76K Non-Thinking
This dataset contains 76,800 fixed teacher trajectories generated for a
prompt-aligned reproduction study of on-policy distillation with Qwen3-1.7B.
It is an independent research artifact, not an official release from the model
or paper authors.
## Models and generation
- Teacher: `Qwen/Qwen3-4B-Instruct-2507`
- Tokenizer/chat template: `Qwen/Qwen3-1.7B`
- Mode: non-thinking (`enable_thinking=False`)
- Temperature: 0.7
- Top-p: 1.0
- Top-k: 0
- Maximum prompt length: 4,096 tokens
- Maximum completion length: 8,192 tokens
- Seed for row `i`: `42 + i`
Prompts were passed to the vLLM completions API as exact token IDs with no
additional special tokens. Returned token IDs and special tokens are retained.
## Statistics
| Statistic | Value |
| --- | ---: |
| Rows | 76,800 |
| Prompt tokens | 21,596,089 |
| Completion tokens | 264,338,320 |
| Total stored tokens | 285,934,409 |
| Mean completion length | 3,441.91 |
| Completions reaching 8,192 tokens | 10,974 (14.29%) |
The first 25,600 trajectories exactly reuse the earlier 100-step teacher bank.
The remaining 51,200 rows extend the same seed-42 prompt permutation without an
epoch wrap.
## Fields
- `input_ids`: concatenated prompt and completion token IDs
- `labels`: `-100` over prompt tokens and completion token IDs thereafter
- `prompt_ids`, `completion_ids`: separate exact token sequences
- `messages`: original conversation plus decoded teacher assistant response
- `domain`: math, science, or code
- `source_dataset`, `source_config`, `source_split`, `source_shard`, `source_id`:
upstream provenance
- `prompt_sha256`: hash of the source prompt
- `source_row_index`, `unified_order_index`: source and rollout schedule indices
- `finish_reason`, `generation_seed`: generation metadata
## Integrity
- Prompt order SHA-256:
`c1cc096fa94f82bf44457673b112039f604aaba19711a850f368514da6d7429d`
- Completion IDs SHA-256:
`a3c5ae8551b23025bb57d6d60a011ef7e156d034d0853656dbb33d29b064127b`
The repository includes the full generation manifest. A post-generation audit
found zero prompt-identity mismatches, and the prompt-token hash matches a fresh
application of the documented Qwen non-thinking chat template.
## Usage
```python
from datasets import load_dataset
dataset = load_dataset(
"YangyiH/qwen3-4b-teacher-rollouts-76k-nonthinking",
split="train",
)
```
## Source data and licensing
Prompts originate from the NVIDIA OpenMathReasoning,
OpenScienceReasoning-2, and OpenCodeReasoning datasets. This repository does
not assert a new blanket license over upstream prompt content. Users should
review and comply with the source datasets' licenses, terms, and attribution
requirements, as well as the Qwen model license.
## Limitations
- The source mixture intentionally retains duplicate prompts inherited from
the upstream datasets.
- No additional quality, difficulty, benchmark-contamination, or safety
filtering was applied.
- 14.29% of completions reached the configured 8,192-token cap and may be
truncated rather than naturally terminated.
- Generated responses can contain errors or undesirable content inherited from
the model and source prompts.
|