vfleaking commited on
Commit
0b95c2d
·
verified ·
1 Parent(s): 5fb353a

Publish CapySFT 500M, 1B and 2B using four shared shards

Browse files
LICENSE.md ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Upstream licenses and attribution
2
+
3
+ CapySFT is derived from multiple sources. This notice does not replace
4
+ their licenses or grant broader rights to their contents.
5
+
6
+ - **Tulu 3 SFT Mixture**: https://huggingface.co/datasets/allenai/tulu-3-sft-mixture
7
+ identifies its collection license as Open Data Commons Attribution 1.0
8
+ (https://opendatacommons.org/licenses/by/1-0/). Its card states that
9
+ component licenses differ and some portions are non-commercial. For example,
10
+ No Robots is identified as CC-BY-NC-4.0. Follow the component links and
11
+ attribution/citation requirements in the upstream card.
12
+ - **SmolTalk**: https://huggingface.co/datasets/HuggingFaceTB/smoltalk
13
+ licenses the newly generated Smol-Magpie-Ultra, Smol-constraints,
14
+ Smol-rewrite, and Smol-summarize datasets under Apache 2.0
15
+ (https://www.apache.org/licenses/LICENSE-2.0). Its existing public
16
+ components retain their original licenses, linked from the upstream
17
+ dataset-composition table.
18
+ - Assistant responses were generated with **DeepSeek-V4-Flash-0731**.
19
+ Source prompt and model-output conditions remain applicable to their
20
+ respective material.
21
+ - The included **Qwen3.5-32K tokenizer** is derived from the Qwen3.5
22
+ tokenizer by retaining the first 32,768 base token IDs and remapping 33
23
+ special tokens. Qwen3.5 provenance is separate from the data mixture.
24
+
25
+ Each record retains `dataset_origin` and `source` labels. The release manifest
26
+ enumerates the labels present in this release. Regenerated responses are
27
+ modifications to the source conversations, and do not imply that the
28
+ upstream datasets endorse this release. No blanket Apache, MIT, or other
29
+ single permissive license is applied to the complete mixture here.
30
+
31
+ Upstream license statements were checked on 2026-11-10 against dataset-card
32
+ revisions `b14afda60f1bbebe55d5d2fa1e4df5042f97f8be` (Tulu 3) and
33
+ `5feaf2fd3ffca7c237fc38d1861bc30365d48ffa` (SmolTalk). These are the
34
+ revisions inspected for this notice, not asserted download revisions of the
35
+ original generation inputs.
README.md ADDED
@@ -0,0 +1,197 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ pretty_name: CapySFT
3
+ task_categories:
4
+ - text-generation
5
+ size_categories:
6
+ - 100K<n<1M
7
+ tags:
8
+ - sft
9
+ - reasoning
10
+ - tool-use
11
+ - synthetic
12
+ source_datasets:
13
+ - allenai/tulu-3-sft-mixture
14
+ - HuggingFaceTB/smoltalk
15
+ configs:
16
+ - config_name: 500M
17
+ data_files:
18
+ - split: train
19
+ path: data/train-00000-of-00004.jsonl.gz
20
+ - config_name: 1B
21
+ default: true
22
+ data_files:
23
+ - split: train
24
+ path:
25
+ - data/train-00000-of-00004.jsonl.gz
26
+ - data/train-00001-of-00004.jsonl.gz
27
+ - config_name: 2B
28
+ data_files:
29
+ - split: train
30
+ path: data/train-*.jsonl.gz
31
+ ---
32
+
33
+ # CapySFT
34
+
35
+ CapySFT is a dataset for supervised fine tuning of assistants that can reason,
36
+ answer questions, and use terminal tools. User requests come from Tulu 3 and
37
+ SmolTalk. New assistant responses were generated with DeepSeek-V4-Flash-0731.
38
+
39
+ ## Dataset sizes
40
+
41
+ Choose `500M`, `1B`, or `2B` supervised response tokens. Token counts use
42
+ our LLM course's truncated Qwen3.5 tokenizer (32,768 base tokens plus 33
43
+ special tokens), included under `tokenizer/`. Counts vary with the tokenizer
44
+ and chat formatting.
45
+
46
+ | Configuration | Source conversations | Training records | Supervised response tokens | Files |
47
+ | --- | ---: | ---: | ---: | ---: |
48
+ | `500M` | 199,455 | 217,998 | 500,024,135 | first 1 |
49
+ | `1B` (default) | 399,746 | 436,342 | 1,000,019,252 | first 2 |
50
+ | `2B` | 785,488 | 862,479 | 1,999,896,726 | all 4 |
51
+
52
+ The configurations share four compressed JSONL files, each with approximately
53
+ 500 million supervised response tokens. Smaller configurations are subsets
54
+ of the larger ones. Related records from the same source conversation stay
55
+ in one file. Each configuration contains only a `train` split.
56
+
57
+ One record is a complete training segment. Long source conversations can
58
+ produce several records when their history is summarized and generation
59
+ continues. Retained segments are not truncated and fit within 16,384 tokens.
60
+ Supervised counts include selected assistant reasoning, answers, tool calls,
61
+ summaries, and the end marker of each selected message. Context messages and
62
+ tool results do not contribute to these counts.
63
+
64
+ ## Load with Hugging Face
65
+
66
+ Use the standard `datasets` interface:
67
+
68
+ ~~~python
69
+ from datasets import load_dataset
70
+
71
+ data = load_dataset("vfleaking/capysft", "1B", split="train", token=True)
72
+ example = data[0]
73
+ print(example["messages"])
74
+ print(example["supervised_message_indices"])
75
+ ~~~
76
+
77
+ Change `"1B"` to `"500M"` or `"2B"` to choose a size. Add `streaming=True`
78
+ to iterate without building a complete local Arrow cache. While the
79
+ repository is private, log in with an account that has access.
80
+
81
+ ## Record format
82
+
83
+ The three fields needed for training are:
84
+
85
+ | Field | Type | Meaning |
86
+ | --- | --- | --- |
87
+ | `messages` | list of message dictionaries | All messages in conversation order, including those used only as context. |
88
+ | `tools` | JSON string or null | Available tool definitions. Decode with `json.loads` before applying a chat template. |
89
+ | `supervised_message_indices` | list of integers | Indices into `messages` identifying assistant messages to supervise. |
90
+
91
+ Messages have `role` and `content`, and may have `reasoning_content`,
92
+ `tool_calls`, or `tool_call_id`. Each tool call has `id`, `type`, and
93
+ `function`. Its `function.arguments` field is a JSON string.
94
+ Missing optional message fields load as `None`. Remove only those null
95
+ fields, preserving empty strings and lists:
96
+
97
+ ~~~python
98
+ import copy
99
+ import json
100
+
101
+ messages = [
102
+ {key: value for key, value in message.items() if value is not None}
103
+ for message in copy.deepcopy(example["messages"])
104
+ ]
105
+ for message in messages:
106
+ for call in message.get("tool_calls", []):
107
+ call["function"]["arguments"] = json.loads(call["function"]["arguments"])
108
+ tools = json.loads(example["tools"]) if example["tools"] is not None else None
109
+ ~~~
110
+
111
+ Each row also has `id`, a unique record identifier, and `conversation_id`,
112
+ shared by records from the same source conversation. Group by
113
+ `conversation_id` when making a validation split to keep related records
114
+ together.
115
+
116
+ The `metadata` dictionary retains source, generation, segmentation, and
117
+ token statistics:
118
+
119
+ | Metadata field | Meaning |
120
+ | --- | --- |
121
+ | `source` | Prompt mixture: `tulu3` or `smoltalk`. |
122
+ | `subset` | Original subset label within the mixture. |
123
+ | `original_global_index` | Source conversation's index in the prepared prompt collection. |
124
+ | `original_conversation_id` | Original generation identifier, retained for tracing a record to its source. |
125
+ | `generation_status` | Recorded generation outcome. Completed segments are retained even when later generation did not complete. |
126
+ | `segment_index` | Segment order within the source conversation. |
127
+ | `segment_type` | `initial`, `compaction` (summary generation), or `continuation` (generation after a summary). |
128
+ | `segment_end_reason` | Recorded reason this segment ended. |
129
+ | `generated_responses` | Number of newly generated assistant messages; equals the length of `supervised_message_indices`. |
130
+ | `response_tokens` | Number of supervised response tokens. |
131
+ | `full_tokens` | Number of input tokens, including context and the final message's end marker. |
132
+
133
+ Other source metadata, when present, is also retained.
134
+
135
+ ## Tokenization and supervision
136
+
137
+ The supplied chat template preserves reasoning on every assistant turn.
138
+ After preparing `messages` and `tools` as above, render and tokenize with:
139
+
140
+ ~~~python
141
+ from transformers import AutoTokenizer
142
+
143
+ tokenizer = AutoTokenizer.from_pretrained(
144
+ "vfleaking/capysft", subfolder="tokenizer", token=True
145
+ )
146
+ text = tokenizer.apply_chat_template(
147
+ messages, tools=tools, tokenize=False, add_generation_prompt=False
148
+ )
149
+ text = text.removesuffix("\n")
150
+ tokens = tokenizer.encode(text, add_special_tokens=False)
151
+ ~~~
152
+
153
+ Supervise only assistant messages selected by `supervised_message_indices`.
154
+ For each selected message, include its opening `<think>`, reasoning,
155
+ closing `</think>`, answer or complete tool call, and closing `<|im_end|>`.
156
+ Include newlines within that body. Mask the assistant role header,
157
+ the newline after its end marker, all other messages, and padding.
158
+
159
+ Earlier assistant messages can be context only. For example, a generated
160
+ summary is supervised when first produced, then appears as masked history
161
+ in a later record. Keep these earlier messages as input.
162
+
163
+ The final `<|im_end|>` ends the segment. Remove only the following template
164
+ newline and do not append another end token. When packing records together,
165
+ keep attention and target shifting within each record.
166
+
167
+ ## How the data was generated
168
+
169
+ We used a small agent harness developed for our LLM course assignments,
170
+ which we call MiniCapyCode. It runs the conversation loop between the model
171
+ and a sandboxed Linux environment. The model may use bash, read, edit, and
172
+ write tools, receive their results, and continue its response. Tool use is
173
+ optional.
174
+
175
+ Generation used thinking enabled, low reasoning effort, `temperature=1`,
176
+ and `top_p=1`. Later user questions could be rewritten in conversation context.
177
+ When a conversation approached the context limit, the harness could ask
178
+ the model to summarize its history and continue from that summary.
179
+ Summary requests and responses are included as separate training records.
180
+
181
+ The generated responses can contain errors. Some source conversations ended
182
+ before all turns completed, but their completed training segments are
183
+ retained. There is no validation or test split, and the data has not been
184
+ checked for overlap with every benchmark.
185
+
186
+ ## Sources and terms
187
+
188
+ The source prompts come from
189
+ [Tulu 3 SFT Mixture](https://huggingface.co/datasets/allenai/tulu-3-sft-mixture)
190
+ and [SmolTalk](https://huggingface.co/datasets/HuggingFaceTB/smoltalk).
191
+ Each record retains the source and subset labels in `metadata`.
192
+
193
+ Upstream terms differ by subset, and some Tulu 3 subsets restrict commercial
194
+ use. Regenerating responses does not remove conditions on the source
195
+ prompts. See [LICENSE.md](LICENSE.md) and the upstream dataset cards for
196
+ licenses, attribution, and terms governing generated outputs. No single
197
+ permissive license is asserted for the entire mixture.
data/train-00000-of-00004.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:097939079e712994ffd599bc79a28c5d727479c5cea3d76f36de11d0574575a3
3
+ size 1017188186
data/train-00001-of-00004.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c3f7bd0ae569da44796785b777e84b36b848136e33a4a51ef5b8e61b72c23660
3
+ size 1013511728
data/train-00002-of-00004.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ff6ac95bb443b2a64952199f2631a332943494d479f5babefb754bfc3d8cc16f
3
+ size 1019335630
data/train-00003-of-00004.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a483141be7aa2b2727d2c020d27be3c835cb778c3c25a11be6ad9151e44d8d01
3
+ size 1026156194
release-manifest.json ADDED
@@ -0,0 +1,357 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset_name": "capysft",
3
+ "version": "1.3",
4
+ "status": "complete",
5
+ "format": "course-sft-jsonl-gzip",
6
+ "record_fields": [
7
+ "id",
8
+ "conversation_id",
9
+ "messages",
10
+ "tools",
11
+ "supervised_message_indices",
12
+ "metadata"
13
+ ],
14
+ "tools_encoding": "JSON string or null",
15
+ "tool_arguments_encoding": "JSON string",
16
+ "configurations": {
17
+ "500M": {
18
+ "source_conversations": 199455,
19
+ "segments": 217998,
20
+ "response_tokens": 500024135,
21
+ "full_tokens": 919367388,
22
+ "compaction_response_tokens": 15745067,
23
+ "files": [
24
+ "data/train-00000-of-00004.jsonl.gz"
25
+ ]
26
+ },
27
+ "1B": {
28
+ "source_conversations": 399746,
29
+ "segments": 436342,
30
+ "response_tokens": 1000019252,
31
+ "full_tokens": 1834551029,
32
+ "compaction_response_tokens": 30893393,
33
+ "files": [
34
+ "data/train-00000-of-00004.jsonl.gz",
35
+ "data/train-00001-of-00004.jsonl.gz"
36
+ ]
37
+ },
38
+ "2B": {
39
+ "source_conversations": 785488,
40
+ "segments": 862479,
41
+ "response_tokens": 1999896726,
42
+ "full_tokens": 3686627495,
43
+ "compaction_response_tokens": 65008945,
44
+ "files": [
45
+ "data/train-00000-of-00004.jsonl.gz",
46
+ "data/train-00001-of-00004.jsonl.gz",
47
+ "data/train-00002-of-00004.jsonl.gz",
48
+ "data/train-00003-of-00004.jsonl.gz"
49
+ ]
50
+ }
51
+ },
52
+ "canonical_shards": [
53
+ {
54
+ "file": "data/train-00000-of-00004.jsonl.gz",
55
+ "bytes": 1017188186,
56
+ "sha256": "097939079e712994ffd599bc79a28c5d727479c5cea3d76f36de11d0574575a3",
57
+ "source_conversations": 199455,
58
+ "segments": 217998,
59
+ "response_tokens": 500024135,
60
+ "full_tokens": 919367388
61
+ },
62
+ {
63
+ "file": "data/train-00001-of-00004.jsonl.gz",
64
+ "bytes": 1013511728,
65
+ "sha256": "c3f7bd0ae569da44796785b777e84b36b848136e33a4a51ef5b8e61b72c23660",
66
+ "source_conversations": 200291,
67
+ "segments": 218344,
68
+ "response_tokens": 499995117,
69
+ "full_tokens": 915183641
70
+ },
71
+ {
72
+ "file": "data/train-00002-of-00004.jsonl.gz",
73
+ "bytes": 1019335630,
74
+ "sha256": "ff6ac95bb443b2a64952199f2631a332943494d479f5babefb754bfc3d8cc16f",
75
+ "source_conversations": 192723,
76
+ "segments": 212748,
77
+ "response_tokens": 498506594,
78
+ "full_tokens": 922709858
79
+ },
80
+ {
81
+ "file": "data/train-00003-of-00004.jsonl.gz",
82
+ "bytes": 1026156194,
83
+ "sha256": "a483141be7aa2b2727d2c020d27be3c835cb778c3c25a11be6ad9151e44d8d01",
84
+ "source_conversations": 193019,
85
+ "segments": 213389,
86
+ "response_tokens": 501370880,
87
+ "full_tokens": 929366608
88
+ }
89
+ ],
90
+ "tokenizer_identity": {
91
+ "name": "Qwen3.5-32K",
92
+ "native_template_sha256": "a4aee8afcf2e0711942cf848899be66016f8d14a889ff9ede07bca099c28f715",
93
+ "counting_template_sha256": "cbad793417178000a8145de58db885ecf45a8f8380393dbbf24373b52f046ddc",
94
+ "preserve_all_thinking": true,
95
+ "add_generation_prompt": true,
96
+ "enable_thinking": true,
97
+ "tool_arguments": "JSON objects decoded on a copy, invalid arguments rejected",
98
+ "training_counting": {
99
+ "implementation": "sft_all_turn_thinking.tokenize_structural",
100
+ "implementation_sha256": "215312de895de1e79d492db251386985e798222f83eb9eeee239a4f57380029e",
101
+ "preserve_all_thinking": true,
102
+ "add_generation_prompt": false,
103
+ "append_eod": false,
104
+ "terminal_separator": "remove the template newline after the final message im_end",
105
+ "tool_arguments": "record passed unchanged to SFT serialization"
106
+ },
107
+ "libraries": {
108
+ "transformers": "5.12.1",
109
+ "tokenizers": "0.22.2",
110
+ "jinja2": "3.1.6"
111
+ },
112
+ "source_assets_sha256": {
113
+ "derivation_report.json": "f40d6ac30b795ec1e1d428a17bd4d54831c8c8e8b9e14aa2d5dd3825dfc6e480",
114
+ "tokenizer.json": "9113f540313f00317aa6e860d01ea73fbac8b0b3af0f5cbf0c93af047cdf6b4c",
115
+ "tokenizer_config.json": "08496e6fef6bab1359e300797cb8ac230e8a3ee652676eaeb261755ab1e548a7"
116
+ },
117
+ "assets_sha256": {
118
+ "tokenizer.json": "9113f540313f00317aa6e860d01ea73fbac8b0b3af0f5cbf0c93af047cdf6b4c",
119
+ "tokenizer_config.json": "f61c4793105d9508248590587e6e94fba2f64a4bb84f64eaa4925d0466847c5b"
120
+ },
121
+ "chat_template_sha256": "cbad793417178000a8145de58db885ecf45a8f8380393dbbf24373b52f046ddc",
122
+ "vocabulary": {
123
+ "base_tokens": 32768,
124
+ "special_tokens": 33,
125
+ "total_tokens": 32801,
126
+ "derivation": "Truncated Qwen3.5 base vocabulary and remapped special tokens, prepared for our LLM course"
127
+ }
128
+ },
129
+ "by_source": {
130
+ "smoltalk": {
131
+ "segments": 360239,
132
+ "response_tokens": 923363273,
133
+ "full_tokens": 1507389801
134
+ },
135
+ "tulu3": {
136
+ "segments": 502240,
137
+ "response_tokens": 1076533453,
138
+ "full_tokens": 2179237694
139
+ }
140
+ },
141
+ "by_segment_type": {
142
+ "initial": {
143
+ "segments": 781693,
144
+ "response_tokens": 1729331270,
145
+ "full_tokens": 2751736831
146
+ },
147
+ "compaction": {
148
+ "segments": 39074,
149
+ "response_tokens": 65008945,
150
+ "full_tokens": 499457744
151
+ },
152
+ "continuation": {
153
+ "segments": 41712,
154
+ "response_tokens": 205556511,
155
+ "full_tokens": 435432920
156
+ }
157
+ },
158
+ "source_run": "ds0731-20261021-r5",
159
+ "source_manifests": {
160
+ "500M": {
161
+ "dataset_name": "capysft-v1.2-500M",
162
+ "sha256": "d4abf04b9499f242422fc03e59540760363e2e9aec1d2427ea64b5b24d04b4a7"
163
+ },
164
+ "1B": {
165
+ "dataset_name": "capysft-v1.3",
166
+ "sha256": "5def852c6632f02b459584136ed25c15576b2f8575c08556fbb11c5dff6d8db9"
167
+ },
168
+ "2B": {
169
+ "dataset_name": "capysft-v1.2-2B",
170
+ "sha256": "afecad953a1526496695935594fe26fbc2b543b370a2cd2a2c7bbd45ccec486a"
171
+ }
172
+ },
173
+ "revision": {
174
+ "kind": "terminal-message-eos",
175
+ "messages_and_response_targets_unchanged": true,
176
+ "removed_template_newline_and_extra_eod": true,
177
+ "segments_reordered_into_nested_shards": true,
178
+ "original_pack_indexes_omitted": true
179
+ },
180
+ "source_labels": [
181
+ {
182
+ "dataset_origin": "smoltalk",
183
+ "source": "everyday-conversations",
184
+ "segments": 1089
185
+ },
186
+ {
187
+ "dataset_origin": "smoltalk",
188
+ "source": "explore-instruct-rewriting",
189
+ "segments": 14505
190
+ },
191
+ {
192
+ "dataset_origin": "smoltalk",
193
+ "source": "metamathqa-50k",
194
+ "segments": 22641
195
+ },
196
+ {
197
+ "dataset_origin": "smoltalk",
198
+ "source": "self-oss-instruct",
199
+ "segments": 23380
200
+ },
201
+ {
202
+ "dataset_origin": "smoltalk",
203
+ "source": "smol-constraints",
204
+ "segments": 16561
205
+ },
206
+ {
207
+ "dataset_origin": "smoltalk",
208
+ "source": "smol-magpie-ultra",
209
+ "segments": 210895
210
+ },
211
+ {
212
+ "dataset_origin": "smoltalk",
213
+ "source": "smol-rewrite",
214
+ "segments": 25213
215
+ },
216
+ {
217
+ "dataset_origin": "smoltalk",
218
+ "source": "smol-summarize",
219
+ "segments": 45955
220
+ },
221
+ {
222
+ "dataset_origin": "tulu3",
223
+ "source": "ai2-adapt-dev/coconot_converted",
224
+ "segments": 6985
225
+ },
226
+ {
227
+ "dataset_origin": "tulu3",
228
+ "source": "ai2-adapt-dev/evol_codealpaca_heval_decontaminated",
229
+ "segments": 88406
230
+ },
231
+ {
232
+ "dataset_origin": "tulu3",
233
+ "source": "ai2-adapt-dev/no_robots_converted",
234
+ "segments": 5179
235
+ },
236
+ {
237
+ "dataset_origin": "tulu3",
238
+ "source": "ai2-adapt-dev/numinamath_tir_math_decontaminated",
239
+ "segments": 40269
240
+ },
241
+ {
242
+ "dataset_origin": "tulu3",
243
+ "source": "ai2-adapt-dev/oasst1_converted",
244
+ "segments": 4438
245
+ },
246
+ {
247
+ "dataset_origin": "tulu3",
248
+ "source": "ai2-adapt-dev/personahub_code_v2_34999",
249
+ "segments": 21137
250
+ },
251
+ {
252
+ "dataset_origin": "tulu3",
253
+ "source": "ai2-adapt-dev/personahub_ifdata_manual_seed_v3_29980",
254
+ "segments": 18254
255
+ },
256
+ {
257
+ "dataset_origin": "tulu3",
258
+ "source": "ai2-adapt-dev/personahub_math_v5_regen_149960",
259
+ "segments": 94819
260
+ },
261
+ {
262
+ "dataset_origin": "tulu3",
263
+ "source": "ai2-adapt-dev/tulu_v3.9_open_math_2_gsm8k_50k",
264
+ "segments": 30072
265
+ },
266
+ {
267
+ "dataset_origin": "tulu3",
268
+ "source": "ai2-adapt-dev/tulu_v3.9_personahub_math_interm_algebra_20k",
269
+ "segments": 12084
270
+ },
271
+ {
272
+ "dataset_origin": "tulu3",
273
+ "source": "ai2-adapt-dev/tulu_v3.9_sciriff_10k",
274
+ "segments": 6020
275
+ },
276
+ {
277
+ "dataset_origin": "tulu3",
278
+ "source": "ai2-adapt-dev/tulu_v3.9_synthetic_finalresp_wildguardmixtrain_decontaminated_50k",
279
+ "segments": 30331
280
+ },
281
+ {
282
+ "dataset_origin": "tulu3",
283
+ "source": "ai2-adapt-dev/tulu_v3.9_wildchat_100k",
284
+ "segments": 84316
285
+ },
286
+ {
287
+ "dataset_origin": "tulu3",
288
+ "source": "ai2-adapt-dev/tulu_v3.9_wildjailbreak_decontaminated_50k",
289
+ "segments": 29994
290
+ },
291
+ {
292
+ "dataset_origin": "tulu3",
293
+ "source": "allenai/tulu-3-sft-personas-math-grade",
294
+ "segments": 29936
295
+ }
296
+ ],
297
+ "validation": {
298
+ "all_source_compressed_and_uncompressed_sha256_checked": true,
299
+ "all_records_parsed": true,
300
+ "nested_content_sha256_matches": true,
301
+ "all_record_counts_and_token_totals_checked": true,
302
+ "tokenizer_assets_checked": true
303
+ },
304
+ "observed_json_types": {
305
+ "arguments. ": [
306
+ "str"
307
+ ],
308
+ "arguments.command": [
309
+ "str"
310
+ ],
311
+ "arguments.content": [
312
+ "str"
313
+ ],
314
+ "arguments.edit_safe": [
315
+ "bool"
316
+ ],
317
+ "arguments.fieldName": [
318
+ "str"
319
+ ],
320
+ "arguments.filePath": [
321
+ "str"
322
+ ],
323
+ "arguments.file_path": [
324
+ "str"
325
+ ],
326
+ "arguments.limit": [
327
+ "int"
328
+ ],
329
+ "arguments.newString": [
330
+ "str"
331
+ ],
332
+ "arguments.offset": [
333
+ "int"
334
+ ],
335
+ "arguments.olString": [
336
+ "str"
337
+ ],
338
+ "arguments.oldString": [
339
+ "str"
340
+ ],
341
+ "arguments.replaceAll": [
342
+ "bool"
343
+ ],
344
+ "arguments.timeout": [
345
+ "int"
346
+ ],
347
+ "arguments.workdir": [
348
+ "str"
349
+ ],
350
+ "messages.content": [
351
+ "str"
352
+ ],
353
+ "messages.reasoning_content": [
354
+ "str"
355
+ ]
356
+ }
357
+ }
tokenizer/tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer/tokenizer_config.json ADDED
@@ -0,0 +1,305 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": false,
3
+ "added_tokens_decoder": {
4
+ "32768": {
5
+ "content": "<|endoftext|>",
6
+ "single_word": false,
7
+ "lstrip": false,
8
+ "rstrip": false,
9
+ "normalized": false,
10
+ "special": true
11
+ },
12
+ "32769": {
13
+ "content": "<|im_start|>",
14
+ "single_word": false,
15
+ "lstrip": false,
16
+ "rstrip": false,
17
+ "normalized": false,
18
+ "special": true
19
+ },
20
+ "32770": {
21
+ "content": "<|im_end|>",
22
+ "single_word": false,
23
+ "lstrip": false,
24
+ "rstrip": false,
25
+ "normalized": false,
26
+ "special": true
27
+ },
28
+ "32771": {
29
+ "content": "<|object_ref_start|>",
30
+ "single_word": false,
31
+ "lstrip": false,
32
+ "rstrip": false,
33
+ "normalized": false,
34
+ "special": true
35
+ },
36
+ "32772": {
37
+ "content": "<|object_ref_end|>",
38
+ "single_word": false,
39
+ "lstrip": false,
40
+ "rstrip": false,
41
+ "normalized": false,
42
+ "special": true
43
+ },
44
+ "32773": {
45
+ "content": "<|box_start|>",
46
+ "single_word": false,
47
+ "lstrip": false,
48
+ "rstrip": false,
49
+ "normalized": false,
50
+ "special": true
51
+ },
52
+ "32774": {
53
+ "content": "<|box_end|>",
54
+ "single_word": false,
55
+ "lstrip": false,
56
+ "rstrip": false,
57
+ "normalized": false,
58
+ "special": true
59
+ },
60
+ "32775": {
61
+ "content": "<|quad_start|>",
62
+ "single_word": false,
63
+ "lstrip": false,
64
+ "rstrip": false,
65
+ "normalized": false,
66
+ "special": true
67
+ },
68
+ "32776": {
69
+ "content": "<|quad_end|>",
70
+ "single_word": false,
71
+ "lstrip": false,
72
+ "rstrip": false,
73
+ "normalized": false,
74
+ "special": true
75
+ },
76
+ "32777": {
77
+ "content": "<|vision_start|>",
78
+ "single_word": false,
79
+ "lstrip": false,
80
+ "rstrip": false,
81
+ "normalized": false,
82
+ "special": true
83
+ },
84
+ "32778": {
85
+ "content": "<|vision_end|>",
86
+ "single_word": false,
87
+ "lstrip": false,
88
+ "rstrip": false,
89
+ "normalized": false,
90
+ "special": true
91
+ },
92
+ "32779": {
93
+ "content": "<|vision_pad|>",
94
+ "single_word": false,
95
+ "lstrip": false,
96
+ "rstrip": false,
97
+ "normalized": false,
98
+ "special": true
99
+ },
100
+ "32780": {
101
+ "content": "<|image_pad|>",
102
+ "single_word": false,
103
+ "lstrip": false,
104
+ "rstrip": false,
105
+ "normalized": false,
106
+ "special": true
107
+ },
108
+ "32781": {
109
+ "content": "<|video_pad|>",
110
+ "single_word": false,
111
+ "lstrip": false,
112
+ "rstrip": false,
113
+ "normalized": false,
114
+ "special": true
115
+ },
116
+ "32782": {
117
+ "content": "<tool_call>",
118
+ "single_word": false,
119
+ "lstrip": false,
120
+ "rstrip": false,
121
+ "normalized": false,
122
+ "special": true
123
+ },
124
+ "32783": {
125
+ "content": "</tool_call>",
126
+ "single_word": false,
127
+ "lstrip": false,
128
+ "rstrip": false,
129
+ "normalized": false,
130
+ "special": true
131
+ },
132
+ "32784": {
133
+ "content": "<|fim_prefix|>",
134
+ "single_word": false,
135
+ "lstrip": false,
136
+ "rstrip": false,
137
+ "normalized": false,
138
+ "special": true
139
+ },
140
+ "32785": {
141
+ "content": "<|fim_middle|>",
142
+ "single_word": false,
143
+ "lstrip": false,
144
+ "rstrip": false,
145
+ "normalized": false,
146
+ "special": true
147
+ },
148
+ "32786": {
149
+ "content": "<|fim_suffix|>",
150
+ "single_word": false,
151
+ "lstrip": false,
152
+ "rstrip": false,
153
+ "normalized": false,
154
+ "special": true
155
+ },
156
+ "32787": {
157
+ "content": "<|fim_pad|>",
158
+ "single_word": false,
159
+ "lstrip": false,
160
+ "rstrip": false,
161
+ "normalized": false,
162
+ "special": true
163
+ },
164
+ "32788": {
165
+ "content": "<|repo_name|>",
166
+ "single_word": false,
167
+ "lstrip": false,
168
+ "rstrip": false,
169
+ "normalized": false,
170
+ "special": true
171
+ },
172
+ "32789": {
173
+ "content": "<|file_sep|>",
174
+ "single_word": false,
175
+ "lstrip": false,
176
+ "rstrip": false,
177
+ "normalized": false,
178
+ "special": true
179
+ },
180
+ "32790": {
181
+ "content": "<tool_response>",
182
+ "single_word": false,
183
+ "lstrip": false,
184
+ "rstrip": false,
185
+ "normalized": false,
186
+ "special": true
187
+ },
188
+ "32791": {
189
+ "content": "</tool_response>",
190
+ "single_word": false,
191
+ "lstrip": false,
192
+ "rstrip": false,
193
+ "normalized": false,
194
+ "special": true
195
+ },
196
+ "32792": {
197
+ "content": "<think>",
198
+ "single_word": false,
199
+ "lstrip": false,
200
+ "rstrip": false,
201
+ "normalized": false,
202
+ "special": true
203
+ },
204
+ "32793": {
205
+ "content": "</think>",
206
+ "single_word": false,
207
+ "lstrip": false,
208
+ "rstrip": false,
209
+ "normalized": false,
210
+ "special": true
211
+ },
212
+ "32794": {
213
+ "content": "<|audio_start|>",
214
+ "single_word": false,
215
+ "lstrip": false,
216
+ "rstrip": false,
217
+ "normalized": false,
218
+ "special": true
219
+ },
220
+ "32795": {
221
+ "content": "<|audio_end|>",
222
+ "single_word": false,
223
+ "lstrip": false,
224
+ "rstrip": false,
225
+ "normalized": false,
226
+ "special": true
227
+ },
228
+ "32796": {
229
+ "content": "<tts_pad>",
230
+ "single_word": false,
231
+ "lstrip": false,
232
+ "rstrip": false,
233
+ "normalized": false,
234
+ "special": true
235
+ },
236
+ "32797": {
237
+ "content": "<tts_text_bos>",
238
+ "single_word": false,
239
+ "lstrip": false,
240
+ "rstrip": false,
241
+ "normalized": false,
242
+ "special": true
243
+ },
244
+ "32798": {
245
+ "content": "<tts_text_eod>",
246
+ "single_word": false,
247
+ "lstrip": false,
248
+ "rstrip": false,
249
+ "normalized": false,
250
+ "special": true
251
+ },
252
+ "32799": {
253
+ "content": "<tts_text_bos_single>",
254
+ "single_word": false,
255
+ "lstrip": false,
256
+ "rstrip": false,
257
+ "normalized": false,
258
+ "special": true
259
+ },
260
+ "32800": {
261
+ "content": "<|audio_pad|>",
262
+ "single_word": false,
263
+ "lstrip": false,
264
+ "rstrip": false,
265
+ "normalized": false,
266
+ "special": true
267
+ }
268
+ },
269
+ "additional_special_tokens": [
270
+ "<|im_start|>",
271
+ "<|im_end|>",
272
+ "<|object_ref_start|>",
273
+ "<|object_ref_end|>",
274
+ "<|box_start|>",
275
+ "<|box_end|>",
276
+ "<|quad_start|>",
277
+ "<|quad_end|>",
278
+ "<|vision_start|>",
279
+ "<|vision_end|>",
280
+ "<|vision_pad|>",
281
+ "<|image_pad|>",
282
+ "<|video_pad|>"
283
+ ],
284
+ "bos_token": null,
285
+ "chat_template": "{%- set image_count = namespace(value=0) %}\n{%- set video_count = namespace(value=0) %}\n{%- macro render_content(content, do_vision_count, is_system_content=false) %}\n {%- if content is string %}\n {{- content }}\n {%- elif content is iterable and content is not mapping %}\n {%- for item in content %}\n {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}\n {%- if is_system_content %}\n {{- raise_exception('System message cannot contain images.') }}\n {%- endif %}\n {%- if do_vision_count %}\n {%- set image_count.value = image_count.value + 1 %}\n {%- endif %}\n {%- if add_vision_id %}\n {{- 'Picture ' ~ image_count.value ~ ': ' }}\n {%- endif %}\n {{- '<|vision_start|><|image_pad|><|vision_end|>' }}\n {%- elif 'video' in item or item.type == 'video' %}\n {%- if is_system_content %}\n {{- raise_exception('System message cannot contain videos.') }}\n {%- endif %}\n {%- if do_vision_count %}\n {%- set video_count.value = video_count.value + 1 %}\n {%- endif %}\n {%- if add_vision_id %}\n {{- 'Video ' ~ video_count.value ~ ': ' }}\n {%- endif %}\n {{- '<|vision_start|><|video_pad|><|vision_end|>' }}\n {%- elif 'text' in item %}\n {{- item.text }}\n {%- else %}\n {{- raise_exception('Unexpected item type in content.') }}\n {%- endif %}\n {%- endfor %}\n {%- elif content is none or content is undefined %}\n {{- '' }}\n {%- else %}\n {{- raise_exception('Unexpected content type.') }}\n {%- endif %}\n{%- endmacro %}\n{%- if not messages %}\n {{- raise_exception('No messages provided.') }}\n{%- endif %}\n{%- if tools and tools is iterable and tools is not mapping %}\n {{- '<|im_start|>system\\n' }}\n {{- \"# Tools\\n\\nYou have access to the following functions:\\n\\n<tools>\" }}\n {%- for tool in tools %}\n {{- \"\\n\" }}\n {{- tool | tojson }}\n {%- endfor %}\n {{- \"\\n</tools>\" }}\n {{- '\\n\\nIf you choose to call a function ONLY reply in the following format with NO suffix:\\n\\n<tool_call>\\n<function=example_function_name>\\n<parameter=example_parameter_1>\\nvalue_1\\n</parameter>\\n<parameter=example_parameter_2>\\nThis is the value for the second parameter\\nthat can span\\nmultiple lines\\n</parameter>\\n</function>\\n</tool_call>\\n\\n<IMPORTANT>\\nReminder:\\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\\n- Required parameters MUST be specified\\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\\n</IMPORTANT>' }}\n {%- if messages[0].role == 'system' %}\n {%- set content = render_content(messages[0].content, false, true)|trim %}\n {%- if content %}\n {{- '\\n\\n' + content }}\n {%- endif %}\n {%- endif %}\n {{- '<|im_end|>\\n' }}\n{%- else %}\n {%- if messages[0].role == 'system' %}\n {%- set content = render_content(messages[0].content, false, true)|trim %}\n {{- '<|im_start|>system\\n' + content + '<|im_end|>\\n' }}\n {%- endif %}\n{%- endif %}\n{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}\n{%- for message in messages[::-1] %}\n {%- set index = (messages|length - 1) - loop.index0 %}\n {%- if ns.multi_step_tool and message.role == \"user\" %}\n {%- set content = render_content(message.content, false)|trim %}\n {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}\n {%- set ns.multi_step_tool = false %}\n {%- set ns.last_query_index = index %}\n {%- endif %}\n {%- endif %}\n{%- endfor %}\n{%- if ns.multi_step_tool %}\n {{- raise_exception('No user query found in messages.') }}\n{%- endif %}\n{%- for message in messages %}\n {%- set content = render_content(message.content, true)|trim %}\n {%- if message.role == \"system\" %}\n {%- if not loop.first %}\n {{- raise_exception('System message must be at the beginning.') }}\n {%- endif %}\n {%- elif message.role == \"user\" %}\n {{- '<|im_start|>' + message.role + '\\n' + content + '<|im_end|>' + '\\n' }}\n {%- elif message.role == \"assistant\" %}\n {%- set reasoning_content = '' %}\n {%- if message.reasoning_content is string %}\n {%- set reasoning_content = message.reasoning_content %}\n {%- else %}\n {%- if '</think>' in content %}\n {%- set reasoning_content = content.split('</think>')[0].rstrip('\\n').split('<think>')[-1].lstrip('\\n') %}\n {%- set content = content.split('</think>')[-1].lstrip('\\n') %}\n {%- endif %}\n {%- endif %}\n {%- set reasoning_content = reasoning_content|trim %}\n {%- if true %}\n {{- '<|im_start|>' + message.role + '\\n<think>\\n' + reasoning_content + '\\n</think>\\n\\n' + content }}\n {%- else %}\n {{- '<|im_start|>' + message.role + '\\n' + content }}\n {%- endif %}\n {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}\n {%- for tool_call in message.tool_calls %}\n {%- if tool_call.function is defined %}\n {%- set tool_call = tool_call.function %}\n {%- endif %}\n {%- if loop.first %}\n {%- if content|trim %}\n {{- '\\n\\n<tool_call>\\n<function=' + tool_call.name + '>\\n' }}\n {%- else %}\n {{- '<tool_call>\\n<function=' + tool_call.name + '>\\n' }}\n {%- endif %}\n {%- else %}\n {{- '\\n<tool_call>\\n<function=' + tool_call.name + '>\\n' }}\n {%- endif %}\n {%- if tool_call.arguments is defined %}\n {%- for args_name, args_value in tool_call.arguments|items %}\n {{- '<parameter=' + args_name + '>\\n' }}\n {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}\n {{- args_value }}\n {{- '\\n</parameter>\\n' }}\n {%- endfor %}\n {%- endif %}\n {{- '</function>\\n</tool_call>' }}\n {%- endfor %}\n {%- endif %}\n {{- '<|im_end|>\\n' }}\n {%- elif message.role == \"tool\" %}\n {%- if loop.previtem and loop.previtem.role != \"tool\" %}\n {{- '<|im_start|>user' }}\n {%- endif %}\n {{- '\\n<tool_response>\\n' }}\n {{- content }}\n {{- '\\n</tool_response>' }}\n {%- if not loop.last and loop.nextitem.role != \"tool\" %}\n {{- '<|im_end|>\\n' }}\n {%- elif loop.last %}\n {{- '<|im_end|>\\n' }}\n {%- endif %}\n {%- else %}\n {{- raise_exception('Unexpected message role.') }}\n {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n {{- '<|im_start|>assistant\\n' }}\n {%- if enable_thinking is defined and enable_thinking is false %}\n {{- '<think>\\n\\n</think>\\n\\n' }}\n {%- else %}\n {{- '<think>\\n' }}\n {%- endif %}\n{%- endif %}",
286
+ "clean_up_tokenization_spaces": false,
287
+ "eos_token": "<|im_end|>",
288
+ "errors": "replace",
289
+ "model_max_length": 262144,
290
+ "pad_token": "<|endoftext|>",
291
+ "split_special_tokens": false,
292
+ "tokenizer_class": "Qwen2Tokenizer",
293
+ "unk_token": null,
294
+ "add_bos_token": false,
295
+ "pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
296
+ "extra_special_tokens": {
297
+ "audio_bos_token": "<|audio_start|>",
298
+ "audio_eos_token": "<|audio_end|>",
299
+ "audio_token": "<|audio_pad|>",
300
+ "image_token": "<|image_pad|>",
301
+ "video_token": "<|video_pad|>",
302
+ "vision_bos_token": "<|vision_start|>",
303
+ "vision_eos_token": "<|vision_end|>"
304
+ }
305
+ }