Jaidchen commited on
Commit
3e9c72b
·
verified ·
1 Parent(s): b2ff298

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +39 -10
README.md CHANGED
@@ -68,18 +68,18 @@ repackaged [Qwen 3.6 27B](https://huggingface.co/Qwen/Qwen3.6-27B) with a simpli
68
  </tr>
69
  <tr>
70
  <td>parameters</td>
71
- <td>27 781 427 952</td>
72
- <td>27 320 697 856</td>
73
  </tr>
74
  <tr style='opacity: 50%'>
75
  <td>vocabulary size</td>
76
- <td>248 320</td>
77
- <td>248 320</td>
78
  </tr>
79
  <tr style='opacity: 50%'>
80
  <td>context size</td>
81
- <td>262 144</td>
82
- <td>262 144</td>
83
  </tr>
84
  <tr style='opacity: 50%'>
85
  <td>MTP</td>
@@ -102,7 +102,10 @@ repackaged [Qwen 3.6 27B](https://huggingface.co/Qwen/Qwen3.6-27B) with a simpli
102
  </td>
103
  <td>
104
  <div style='font-family: Jaidevka Code, JetBrains Mono, monospace; line-height: initial'>
105
- <span style='color: hsl(from currentColor 0 100% l)'>do_sample</span>: <span style='color: hsl(from currentColor 50 80% l)'>false</span>
 
 
 
106
  </div>
107
  </td>
108
  </tr>
@@ -129,7 +132,7 @@ repackaged [Qwen 3.6 27B](https://huggingface.co/Qwen/Qwen3.6-27B) with a simpli
129
  <tr>
130
  <td>Jinja template</td>
131
  <td><a href='https://huggingface.co/Qwen/Qwen3.5-27B/blob/main/chat_template.jinja'>Qwen original</a></td>
132
- <td><a href='https://huggingface.co/Qwen/Qwen3.5-27B/blob/main/chat_template.jinja'>Qwen original</a> + <a href='https://huggingface.co/unsloth/Qwen3.5-27B/blob/main/chat_template.jinja'>Unsloth tweaks</a> + custom unimodality patch</td>
133
  </tr>
134
  </table>
135
 
@@ -147,14 +150,40 @@ repackaged [Qwen 3.6 27B](https://huggingface.co/Qwen/Qwen3.6-27B) with a simpli
147
  - Pictures and video frames can still be present in the context without crashing, but their contents are no longer interpreted by the model and won’t do anything else than waste space.
148
  - If you occasionally rely on those capabilities, I suggest adding a `consult` tool to your harness that calls a vision-enabled subagent model like [Gemini Flash](https://openrouter.ai/~google/gemini-flash-latest) or [GPT](https://openrouter.ai/~openai/gpt-latest).
149
  - reduced compatibility
150
- - The applied coercions may confuse your inference engine in case it has fixed expectations about the model’s architecture and thus lead to unpredictable behavior.
151
  - The simplified architecture is handled by the `Qwen3_5ForCausalLM` class which may not be included in your inference engine. In this case you would need to ask your agent or integrate it yourself.
 
152
 
153
  ## caveats
154
 
155
  - model file not split, possibly causing issues if intended to be stored on an HDD from the previous century
156
  - random sampling disabled by default, less suitable for long-form writing, entertainment and casual chat
157
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
158
  ## license
159
 
160
- Apache 2.0, derived from [Qwen 3.6 27B](https://huggingface.co/Qwen/Qwen3.5-27B/blob/main/LICENSE)
 
68
  </tr>
69
  <tr>
70
  <td>parameters</td>
71
+ <td>27 781 427 952</td>
72
+ <td>27 320 697 856</td>
73
  </tr>
74
  <tr style='opacity: 50%'>
75
  <td>vocabulary size</td>
76
+ <td>248 320</td>
77
+ <td>248 320</td>
78
  </tr>
79
  <tr style='opacity: 50%'>
80
  <td>context size</td>
81
+ <td>262 144</td>
82
+ <td>262 144</td>
83
  </tr>
84
  <tr style='opacity: 50%'>
85
  <td>MTP</td>
 
102
  </td>
103
  <td>
104
  <div style='font-family: Jaidevka Code, JetBrains Mono, monospace; line-height: initial'>
105
+ <span style='color: hsl(from currentColor 0 100% l)'>do_sample</span>: <span style='color: hsl(from currentColor 50 80% l)'>false</span><br>
106
+ <span style='color: hsl(from currentColor 0 100% l)'>temperature</span>: <span style='color: hsl(from currentColor 50 80% l)'>0</span><br>
107
+ <span style='color: hsl(from currentColor 0 100% l)'>top_k</span>: <span style='color: hsl(from currentColor 50 80% l)'>1</span><br>
108
+ <span style='color: hsl(from currentColor 0 100% l)'>top_p</span>: <span style='color: hsl(from currentColor 50 80% l)'>1</span></div>
109
  </div>
110
  </td>
111
  </tr>
 
132
  <tr>
133
  <td>Jinja template</td>
134
  <td><a href='https://huggingface.co/Qwen/Qwen3.5-27B/blob/main/chat_template.jinja'>Qwen original</a></td>
135
+ <td><a href='https://huggingface.co/Qwen/Qwen3.5-27B/blob/main/chat_template.jinja'>Qwen original</a> + <a href='https://huggingface.co/unsloth/Qwen3.5-27B/blob/main/chat_template.jinja'>Unsloth tweaks</a> + <a href='https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates/blob/main/chat_template.jinja'>Froggeric tweaks</a> + unimodality patch + further custom tweaks</td>
136
  </tr>
137
  </table>
138
 
 
150
  - Pictures and video frames can still be present in the context without crashing, but their contents are no longer interpreted by the model and won’t do anything else than waste space.
151
  - If you occasionally rely on those capabilities, I suggest adding a `consult` tool to your harness that calls a vision-enabled subagent model like [Gemini Flash](https://openrouter.ai/~google/gemini-flash-latest) or [GPT](https://openrouter.ai/~openai/gpt-latest).
152
  - reduced compatibility
 
153
  - The simplified architecture is handled by the `Qwen3_5ForCausalLM` class which may not be included in your inference engine. In this case you would need to ask your agent or integrate it yourself.
154
+ - The applied coercions may confuse your inference engine in case it has fixed expectations about the model’s architecture and thus lead to unpredictable behavior.
155
 
156
  ## caveats
157
 
158
  - model file not split, possibly causing issues if intended to be stored on an HDD from the previous century
159
  - random sampling disabled by default, less suitable for long-form writing, entertainment and casual chat
160
 
161
+ ## Jinja template changes
162
+
163
+ - base: [Qwen original](https://huggingface.co/Qwen/Qwen3.5-27B/blob/main/chat_template.jinja)
164
+ - adopted tweaks from [Unsloth version](https://huggingface.co/unsloth/Qwen3.5-27B/blob/main/chat_template.jinja)
165
+ - Added `developer` role as alias for `system`. ([line 45–57](https://huggingface.co/unsloth/Qwen3.5-27B/blob/909b2cc7dc1737891ad8f5a7213b702b39410e43/chat_template.jinja#L45-L57))
166
+ - Merged all leading system/developer messages into a single policy message. ([line 58–75](https://huggingface.co/unsloth/Qwen3.5-27B/blob/909b2cc7dc1737891ad8f5a7213b702b39410e43/chat_template.jinja#L58-L75))
167
+ - Allowed histories without a normal human query. ([line 76–87](https://huggingface.co/unsloth/Qwen3.5-27B/blob/909b2cc7dc1737891ad8f5a7213b702b39410e43/chat_template.jinja#L76-L87))
168
+ - Retained later system/developer messages as policy messages. ([line 87–89](https://huggingface.co/unsloth/Qwen3.5-27B/blob/909b2cc7dc1737891ad8f5a7213b702b39410e43/chat_template.jinja#L87-L89), [line 147–149](https://huggingface.co/unsloth/Qwen3.5-27B/blob/909b2cc7dc1737891ad8f5a7213b702b39410e43/chat_template.jinja#L147-L149))
169
+ - Made mapping-argument rendering portable by avoiding `|items`. ([line 108–133](https://huggingface.co/unsloth/Qwen3.5-27B/blob/909b2cc7dc1737891ad8f5a7213b702b39410e43/chat_template.jinja#L108-L133), [line 122–130](https://huggingface.co/unsloth/Qwen3.5-27B/blob/909b2cc7dc1737891ad8f5a7213b702b39410e43/chat_template.jinja#L122-L130))
170
+ - adopted tweaks from [Froggeric version](https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates/blob/b225c1d63f84c65af7dd6af2ba384eb037aa885e/chat_template.jinja)
171
+ - Added `preserve_thinking` option to retain historical reasoning. ([line 8](https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates/blob/b225c1d63f84c65af7dd6af2ba384eb037aa885e/chat_template.jinja#L8), [line 225–230](https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates/blob/b225c1d63f84c65af7dd6af2ba384eb037aa885e/chat_template.jinja#L225-L230))
172
+ - Retained mid-conversation system/developer messages as Qwen system turns. ([line 165–182](https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates/blob/b225c1d63f84c65af7dd6af2ba384eb037aa885e/chat_template.jinja#L165-L182), [line 167–180](https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates/blob/b225c1d63f84c65af7dd6af2ba384eb037aa885e/chat_template.jinja#L167-L180))
173
+ - Allowed string-valued `message.thinking` as fallback for historical reasoning. ([line 183–197](https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates/blob/b225c1d63f84c65af7dd6af2ba384eb037aa885e/chat_template.jinja#L183-L197))
174
+ - Applied boundary-aware `</think>` parsing. ([line 198–223](https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates/blob/b225c1d63f84c65af7dd6af2ba384eb037aa885e/chat_template.jinja#L198-L223))
175
+ - Avoided synthesizing empty historical thinking blocks. ([line 225–230](https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates/blob/b225c1d63f84c65af7dd6af2ba384eb037aa885e/chat_template.jinja#L225-L230))
176
+ - Supported both wrapped and direct tool calls, but never null-wrapped. ([line 231–237](https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates/blob/b225c1d63f84c65af7dd6af2ba384eb037aa885e/chat_template.jinja#L231-L237))
177
+ - Preserved non-empty string tool arguments. ([line 261–280](https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates/blob/b225c1d63f84c65af7dd6af2ba384eb037aa885e/chat_template.jinja#L261-L280), [line 277–279](https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates/blob/b225c1d63f84c65af7dd6af2ba384eb037aa885e/chat_template.jinja#L277-L279))
178
+ - Used direct message indexing for tool-response grouping instead of `loop.previtem`/`loop.nextitem`. ([line 165](https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates/blob/b225c1d63f84c65af7dd6af2ba384eb037aa885e/chat_template.jinja#L165), [line 286–318](https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates/blob/b225c1d63f84c65af7dd6af2ba384eb037aa885e/chat_template.jinja#L286-L318))
179
+ - custom tweaks
180
+ - Merged any number of leading system/developer messages, not just two.
181
+ - Applied boundary-aware `</think>` parsing but excluded malformed-tag recovery.
182
+ - Raised an error for unknown roles instead of silently omitting or converting them.
183
+ - Used direct message indexing for tool-response grouping without Froggeric’s error-escalation state.
184
+ - unimodality patch
185
+ - Rendered image and video content as `[image]` and `[video]`.
186
+
187
  ## license
188
 
189
+ Apache 2.0 – inherited from [Qwen 3.6 27B](https://huggingface.co/Qwen/Qwen3.5-27B/blob/main/LICENSE)