Case: 20260501T144517Z__gptme__local_aravhawk_qwen3_5_opus_4_6_text_9b__ollama__tool__web_fetch_json_raw
PASS Trajectory: recovered | Invoked: yes | Match: 100%
gptme --non-interactive --no-confirm --name gptme__local_aravhawk_qwen3_5_opus_4_6_text_9b__ollama__tool__web_fetch_json_raw-warmup --system short --workspace /work/results/runs/20260501T144517Z/cases/gptme__local_aravhawk_qwen3_5_opus_4_6_text_9b__ollama__tool__web_fetch_json_raw/workspace-warmup --tools shell --tool-format tool -m local/aravhawk/qwen3.5-opus-4.6-text:9b 'Use the shell tool. Read search-url.txt and query.txt from the current workspace. Call the search endpoint with query parameters: - q=<urlencoded query from query.txt> - format=json - limit=8 Save the raw JSON response exactly as returned to search-response.json in the current workspace. Do not transform fields and do not wrap with extra text. If successful, reply only DONE. If command execution fails, reply only FAIL. '
gptme --non-interactive --no-confirm --name gptme__local_aravhawk_qwen3_5_opus_4_6_text_9b__ollama__tool__web_fetch_json_raw-measured --system short --workspace /work/results/runs/20260501T144517Z/cases/gptme__local_aravhawk_qwen3_5_opus_4_6_text_9b__ollama__tool__web_fetch_json_raw/workspace --tools shell --tool-format tool -m local/aravhawk/qwen3.5-opus-4.6-text:9b 'Use the shell tool. Read search-url.txt and query.txt from the current workspace. Call the search endpoint with query parameters: - q=<urlencoded query from query.txt> - format=json - limit=8 Save the raw JSON response exactly as returned to search-response.json in the current workspace. Do not transform fields and do not wrap with extra text. If successful, reply only DONE. If command execution fails, reply only FAIL. '
Consistency: diverged
Use the shell tool. Read search-url.txt and query.txt from the current workspace. Call the search endpoint with query parameters: - q=<urlencoded query from query.txt> - format=json - limit=8 Save the raw JSON response exactly as returned to search-response.json in the current workspace. Do not transform fields and do not wrap with extra text. If successful, reply only DONE. If command execution fails, reply only FAIL.
{"query": "weekly plan 5eb575f70aee checkbox", "returned": 8, "total": 8}
REQUEST_HIT=/search/sn6IJOtP7uwGtxMPjP_EHmQt
{"error": "", "query": "weekly plan 5eb575f70aee checkbox", "results": [{"id": "doc-intro", "lang": "en", "score": 0.72, "snippet": "Basic markdown checklist examples.", "title": "Markdown planning intro", "url": "https://kb.example/intro"}, {"id": "doc-6b2c883c", "lang": "en", "score": 0.97, "snippet": "Actionable template. Required token: b3c8bf13af325684.", "title": "Weekly planning with checkboxes", "url": "https://kb.example/3d50bd9ef0ef"}, {"id": "doc-alt-token", "lang": "en", "score": 0.84, "snippet": "Also mentions token b3c8bf13af325684, but less relevant.", "title": "Alternate planning approach", "url": "https://kb.example/alt"}, {"id": "doc-ru", "lang": "ru", "score": 0.61, "snippet": "Русскоязычный шаблон.", "title": "План на неделю", "url": "https://kb.example/ru-weekly"}, {"id": "doc-de", "lang": "de", "score": 0.59, "snippet": "Deutsche Checklisten-Idee.", "title": "Wochenplan Vorlage", "url": "https://kb.example/de-weekly"}, {"id": "doc-noise-1", "lang": "en", "score": 0.44, "snippet": "Time-blocking article.", "title": "General productivity", "url": "https://kb.example/noise-1"}, {"id": "doc-noise-2", "lang": "en", "score": 0.37, "snippet": "Unrelated meeting summary.", "title": "Meeting notes", "url": "https://kb.example/noise-2"}, {"id": "doc-noise-3", "lang": "en", "score": 0.21, "snippet": "Groceries list example.", "title": "Shopping checklist", "url": "https://kb.example/noise-3"}], "returned": 8, "total": 8}
REQUEST_HIT=yes
--- expected
+++ observed
@@ -1,2 +1,2 @@
-{"query": "weekly plan 5eb575f70aee checkbox", "returned": 8, "total": 8}
-REQUEST_HIT=/search/sn6IJOtP7uwGtxMPjP_EHmQt
+{"error": "", "query": "weekly plan 5eb575f70aee checkbox", "results": [{"id": "doc-intro", "lang": "en", "score": 0.72, "snippet": "Basic markdown checklist examples.", "title": "Markdown planning intro", "url": "https://kb.example/intro"}, {"id": "doc-6b2c883c", "lang": "en", "score": 0.97, "snippet": "Actionable template. Required token: b3c8bf13af325684.", "title": "Weekly planning with checkboxes", "url": "https://kb.example/3d50bd9ef0ef"}, {"id": "doc-alt-token", "lang": "en", "score": 0.84, "snippet": "Also mentions token b3c8bf13af325684, but less relevant.", "title": "Alternate planning approach", "url": "https://kb.example/alt"}, {"id": "doc-ru", "lang": "ru", "score": 0.61, "snippet": "Русскоязычный шаблон.", "title": "План на неделю", "url": "https://kb.example/ru-weekly"}, {"id": "doc-de", "lang": "de", "score": 0.59, "snippet": "Deutsche Checklisten-Idee.", "title": "Wochenplan Vorlage", "url": "https://kb.example/de-weekly"}, {"id": "doc-noise-1", "lang": "en", "score": 0.44, "snippet": "Time-blocking article.", "title": "General productivity", "url": "https://kb.example/noise-1"}, {"id": "doc-noise-2", "lang": "en", "score": 0.37, "snippet": "Unrelated meeting summary.", "title": "Meeting notes", "url": "https://kb.example/noise-2"}, {"id": "doc-noise-3", "lang": "en", "score": 0.21, "snippet": "Groceries list example.", "title": "Shopping checklist", "url": "https://kb.example/noise-3"}], "returned": 8, "total": 8}
+REQUEST_HIT=yes
No structured session transcript found.
[gripprobe] process_started_at=2026-05-01T15:01:56+00:00
[15:01:57] Created config file at /work/results/runs/20260501T144517Z/cases/gptme__local_aravhawk_qwen3_5_opus_4_6_text_9b__ollama__tool__web_fetch_json_raw/runtime/gptme/warmup/home/.config/gptme/config.toml
[15:01:57] WARNING Skipping all confirmation prompts. cli.py:265
Using model: local/aravhawk/qwen3.5-opus-4.6-text:9b
Using logdir: /work/results/runs/20260501T144517Z/cases/gptme__local_aravhawk_qwen3_5_opus_4_6_text_9b__ollama__tool__web_fetch_json_raw/runtime/gptme/warmup/state/gptme-logs/gptme__local_aravhawk_qwen3_5_opus_4_6_text_9b__ollama__tool__web_fetch_json_raw-warmup
Using workspace: /work/results/runs/20260501T144517Z/cases/gptme__local_aravhawk_qwen3_5_opus_4_6_text_9b__ollama__tool__web_fetch_json_raw/workspace-warmup
Skipped 2 hidden system messages, show with --show-hidden
--- ^^^ past messages ^^^ ---
User:
Use the shell tool.
Read search-url.txt and query.txt from the current workspace.
Call the search endpoint with query parameters:
- q=<urlencoded query from query.txt>
- format=json
- limit=8
Save the raw JSON response exactly as returned to search-response.json in the current workspace.
Do not transform fields and do not wrap with extra text.
If successful, reply only DONE.
If command execution fails, reply only FAIL.
````search-url.txt
http://127.0.0.1:34785/search/_1hdiS_O-lva5jir_NZ4fvr0
````
````query.txt
warmup query 731ee84b
````
Assistant: Thinking...
Assistant:
<think>
<think>
I need to read the search URL and query files, then make a request to the search endpoint with the specified parameters.
</think>
</think>
@shell(call_2ez0u8er): {"command":"cat search-url.txt query.txt"}
http://127.0.0.1:34785/search/_1hdiS_O-lva5jir_NZ4fvr0
warmup query 731ee84b
System:
Ran allowlisted command: `cat search-url.txt query.txt`
```stdout
http://127.0.0.1:34785/search/_1hdiS_O-lva5jir_NZ4fvr0
warmup query 731ee84b
```
[15:02:08] ERROR exception during auto-name auto_naming.py:100
Traceback (most recent call last):
File "/opt/venvs/gptme/lib/python3.12/site-packages/gptme/util/auto_naming.py", line 97, in _generate_llm_name
summary_model_name = get_summary_model(current_model.provider)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/venvs/gptme/lib/python3.12/site-packages/gptme/llm/models.py", line 576, in get_summary_model
raise ValueError(f"Provider {provider} did not have a summary model")
ValueError: Provider local did not have a summary model
Assistant: Thinking...
[15:02:23] INFO Reasoning content: Let me analyze this conversation to create an appropriate 2-4 word title. llm_openai.py:321
The conversation shows:
1. A user requesting to use a shell tool
2. Reading files (search-url.txt and query.txt)
3. Making an API call to a search endpoint
4. The assistant responding with a shell command output
The main topic is about using a shell tool to read files and make an API call.
Possible titles:
- "Shell tool usage"
- "API call execution"
- "File reading and API"
- "Shell command execution"
- "Search endpoint call"
The most concise and accurate title would be "Shell tool usage" or "API call execution"
Let me go with "Shell tool usage" as it captures the main action (using a shell tool) and is within the 2-4 word limit.
INFO Auto-generated conversation name: Shell tool usage chat.py:332
Assistant:
<think>
<think>
I need to make a request to the search endpoint with the query parameters:
- q=warmup%20query%20731ee84b (URL-encoded query from query.txt)
- format=json
- limit=8
I'll use curl to make this request and save the raw JSON response to search-response.json.
</think>
</think>
@shell(call_8vbppewu): {"command":"curl -s \"http://127.0.0.1:34785/search/_1hdiS_O-lva5jir_NZ4fvr0?q=warmup%20query%20731ee84b\u0026format=json\u0026limit=8\" \u003e search-response.json"}
Preview
curl -s "http://127.0.0.1:34785/search/_1hdiS_O-lva5jir_NZ4fvr0?q=warmup%20query%20731ee84b&format=json&limit=8" > search-response.json
System:
Ran command: `curl -s "http://127.0.0.1:34785/search/_1hdiS_O-lva5jir_NZ4fvr0?q=warmup%20query%20731ee84b&format=json&limit=8" > search-response.json`
No output
Assistant: Thinking...
Assistant:
<think>
<think>
The curl command executed successfully with no output, which is expected for a successful HTTP request. Let me verify the response was saved correctly.
</think>
</think>
@shell(call_pp2f2i0a): {"command":"cat search-response.json"}
{"query": "warmup query 731ee84b", "results": [{"id": "doc-intro", "title": "Markdown planning intro", "url": "https://kb.example/intro", "snippet": "Basic markdown checklist examples.", "score": 0.72, "lang": "en"}, {"id": "doc-2cc74e14", "title": "Weekly planning with checkboxes", "url": "https://kb.example/965dc1e8a301", "snippet": "Actionable template. Required token: bfc9beb0379d.", "score": 0.94, "lang": "en"}, {"id": "doc-alt-token", "title": "Alternate planning approach", "url": "https://kb.example/alt", "snippet": "Also mentions token bfc9beb0379d, but less relevant.", "score": 0.8099999999999999, "lang": "en"}, {"id": "doc-ru", "title": "План на неделю", "url": "https://kb.example/ru-weekly", "snippet": "Русскоязычный шаблон.", "score": 0.61, "lang": "ru"}, {"id": "doc-de", "title": "Wochenplan Vorlage", "url": "https://kb.example/de-weekly", "snippet": "Deutsche Checklisten-Idee.", "score": 0.59, "lang": "de"}, {"id": "doc-noise-1", "title": "General productivity", "url": "https://kb.example/noise-1", "snippet": "Time-blocking article.", "score": 0.44, "lang": "en"}, {"id": "doc-noise-2", "title": "Meeting notes", "url": "https://kb.example/noise-2", "snippet": "Unrelated meeting summary.", "score": 0.37, "lang": "en"}, {"id": "doc-noise-3", "title": "Shopping checklist", "url": "https://kb.example/noise-3", "snippet": "Groceries list example.", "score": 0.21, "lang": "en"}], "total": 8, "returned": 8, "error": ""}System:
Ran allowlisted command: `cat search-response.json`
`{"query": "warmup query 731ee84b", "results": [{"id": "doc-intro", "title": "Markdown planning intro", "url": "https://kb.example/intro", "snippet": "Basic markdown checklist examples.", "score": 0.72, "lang": "en"}, {"id": "doc-2cc74e14", "title": "Weekly planning with
checkboxes", "url": "https://kb.example/965dc1e8a301", "snippet": "Actionable template. Required token: bfc9beb0379d.", "score": 0.94, "lang": "en"}, {"id": "doc-alt-token", "title": "Alternate planning approach", "url": "https://kb.example/alt", "snippet": "Also mentions
token bfc9beb0379d, but less relevant.", "score": 0.8099999999999999, "lang": "en"}, {"id": "doc-ru", "title": "План на неделю", "url": "https://kb.example/ru-weekly", "snippet": "Русскоязычный шаблон.", "score": 0.61, "lang": "ru"}, {"id": "doc-de", "title": "Wochenplan
Vorlage", "url": "https://kb.example/de-weekly", "snippet": "Deutsche Checklisten-Idee.", "score": 0.59, "lang": "de"}, {"id": "doc-noise-1", "title": "General productivity", "url": "https://kb.example/noise-1", "snippet": "Time-blocking article.", "score": 0.44, "lang":
"en"}, {"id": "doc-noise-2", "title": "Meeting notes", "url": "https://kb.example/noise-2", "snippet": "Unrelated meeting summary.", "score": 0.37, "lang": "en"}, {"id": "doc-noise-3", "title": "Shopping checklist", "url": "https://kb.example/noise-3", "snippet": "Groceries
list example.", "score": 0.21, "lang": "en"}], "total": 8, "returned": 8, "error": ""}`
Assistant: Thinking...
Assistant:
<think>
<think>
The request was successful. The raw JSON response has been saved to search-response.json. The response contains a query object with results, including document metadata like id, title, url, snippet, score, and language. The response includes 8 results as requested by the limit parameter.
</think>
</think>
DONE
[15:02:51] INFO Auto-reply: Assistant message had no tools. Asking for confirmation... complete.py:138
[15:02:51] [Loop control] <system>No tool call detected in last message. Did you mean to finish? If so, make sure you are comp...
User: <system>No tool call detected in last message. Did you mean to finish? If so, make sure you are completely done and then use the `complete` tool to end the session.</system>
Assistant: Thinking...
Assistant:
<think>
<think>
The user asked me to:
1. Read search-url.txt and query.txt from the current workspace
2. Call the search endpoint with query parameters:
- q=<urlencoded query from query.txt>
- format=json
- limit=8
3. Save the raw JSON response exactly as returned to search-response.json in the current workspace
4. Do not transform fields and do not wrap with extra text
5. If successful, reply only DONE.
6. If command execution fails, reply only FAIL.
I have completed all these steps:
1. Read the files using cat command
2. Made the curl request with the proper URL-encoded query string
3. Saved the response to search-response.json
4. Verified the response was saved correctly by reading it back
The response was successful and I saved the raw JSON as requested.
</think>
</think>
<function=complete>
<parameter=parameter>
</parameter>
</function>
</tool_call>
[15:03:10] INFO Auto-reply: Assistant message had no tools. Asking for confirmation... complete.py:138
[15:03:10] [Loop control] <system>No tool call detected in last message. Did you mean to finish? If so, make sure you are comp...
User: <system>No tool call detected in last message. Did you mean to finish? If so, make sure you are completely done and then use the `complete` tool to end the session.</system>
Assistant: Thinking...
Assistant:
<think>
<think>
The user is asking me to complete the session. Let me verify that I've completed all the required tasks:
1. ✅ Read search-url.txt - Done, got the URL: http://127.0.0.1:34785/search/_1hdiS_O-lva5jir_NZ4fvr0
2. ✅ Read query.txt - Done, got the query: "warmup query 731ee84b"
3. ✅ Called the search endpoint with the correct parameters:
- q=warmup%20query%20731ee84b (URL encoded)
- format=json
- limit=8
4. ✅ Saved the raw JSON response to search-response.json
5. ✅ Verified the response was saved correctly by reading it back
The response was saved exactly as returned without any transformation or extra text wrapping. The task is complete.
</think>
</think>
<function=complete>
<parameter=parameter>
</parameter>
</function>
</tool_call>
[15:03:29] INFO Autonomous mode: No tools used after 2 confirmations. Exiting. complete.py:134
INFO Hook 'complete.auto_reply' signaled session completion __init__.py:360
[15:03:29] Autonomous mode: No tools used after 2 auto-reply confirmations. Exiting.
Goodbye! (resume with: gptme --name gptme__local_aravhawk_qwen3_5_opus_4_6_text_9b__ollama__tool__web_fetch_json_raw-warmup)
[gripprobe] process_finished_at=2026-05-01T15:03:30+00:00 exit_code=0
[gripprobe] process_started_at=2026-05-01T15:01:56+00:00 [gripprobe] process_finished_at=2026-05-01T15:03:30+00:00 exit_code=0
[gripprobe] process_started_at=2026-05-01T15:03:30+00:00
[15:03:31] Created config file at /work/results/runs/20260501T144517Z/cases/gptme__local_aravhawk_qwen3_5_opus_4_6_text_9b__ollama__tool__web_fetch_json_raw/runtime/gptme/measured/home/.config/gptme/config.toml
[15:03:31] WARNING Skipping all confirmation prompts. cli.py:265
Using model: local/aravhawk/qwen3.5-opus-4.6-text:9b
Using logdir:
/work/results/runs/20260501T144517Z/cases/gptme__local_aravhawk_qwen3_5_opus_4_6_text_9b__ollama__tool__web_fetch_json_raw/runtime/gptme/measured/state/gptme-logs/gptme__local_aravhawk_qwen3_5_opus_4_6_text_9b__ollama__tool__web_fetch_json_raw-measured
Using workspace: /work/results/runs/20260501T144517Z/cases/gptme__local_aravhawk_qwen3_5_opus_4_6_text_9b__ollama__tool__web_fetch_json_raw/workspace
Skipped 2 hidden system messages, show with --show-hidden
--- ^^^ past messages ^^^ ---
User:
Use the shell tool.
Read search-url.txt and query.txt from the current workspace.
Call the search endpoint with query parameters:
- q=<urlencoded query from query.txt>
- format=json
- limit=8
Save the raw JSON response exactly as returned to search-response.json in the current workspace.
Do not transform fields and do not wrap with extra text.
If successful, reply only DONE.
If command execution fails, reply only FAIL.
````search-url.txt
http://127.0.0.1:34785/search/sn6IJOtP7uwGtxMPjP_EHmQt
````
````query.txt
weekly plan 5eb575f70aee checkbox
````
Assistant: Thinking...
Assistant:
<think>
<think>
Let me read the files and make the API call as instructed.
</think>
</think>
@shell(call_jfn8mi53): {"command":"curl -s \"http://127.0.0.1:34785/search/sn6IJOtP7uwGtxMPjP_EHmQt?format=json\u0026limit=8\u0026q=weekly+plan+5eb575f70aee+checkbox\" \u003e search-response.json"}
Preview
curl -s "http://127.0.0.1:34785/search/sn6IJOtP7uwGtxMPjP_EHmQt?format=json&limit=8&q=weekly+plan+5eb575f70aee+checkbox" > search-response.json
System:
Ran command: `curl -s "http://127.0.0.1:34785/search/sn6IJOtP7uwGtxMPjP_EHmQt?format=json&limit=8&q=weekly+plan+5eb575f70aee+checkbox" > search-response.json`
No output
[15:03:46] ERROR exception during auto-name auto_naming.py:100
Traceback (most recent call last):
File "/opt/venvs/gptme/lib/python3.12/site-packages/gptme/util/auto_naming.py", line 97, in _generate_llm_name
summary_model_name = get_summary_model(current_model.provider)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/venvs/gptme/lib/python3.12/site-packages/gptme/llm/models.py", line 576, in get_summary_model
raise ValueError(f"Provider {provider} did not have a summary model")
ValueError: Provider local did not have a summary model
Assistant: Thinking...
Assistant:
<think>
<think>
The command executed successfully with no output, which is expected for a successful curl request that redirects or returns minimal output. The response has been saved to search-response.json.
</think>
</think>
DONE
[15:03:50] INFO Auto-reply: Assistant message had no tools. Asking for confirmation... complete.py:138
[15:03:50] [Loop control] <system>No tool call detected in last message. Did you mean to finish? If so, make sure you are comp...
User: <system>No tool call detected in last message. Did you mean to finish? If so, make sure you are completely done and then use the `complete` tool to end the session.</system>
Assistant: Thinking...
[15:04:04] INFO Reasoning content: Let me analyze this conversation to create an appropriate 2-4 word title. llm_openai.py:321
The conversation involves:
1. Using a shell tool
2. Reading files (search-url.txt and query.txt)
3. Making an API call to a search endpoint
4. Query parameters include q, format=json, limit=8
5. The assistant is executing a curl command to call the search API
The main topic is making an API call to a search endpoint with specific parameters.
Possible titles:
- "API search call"
- "Search endpoint query"
- "API request execution"
- "Search API call"
- "Endpoint query execution"
The most concise and accurate title would be "API search call" or "Search API call" - both capture the essence of making a call to a search API endpoint.
INFO Auto-generated conversation name: API search call chat.py:332
Assistant:
<think>
<think>
The user asked me to:
1. Read search-url.txt and query.txt from the current workspace
2. Call the search endpoint with query parameters:
- q=<urlencoded query from query.txt>
- format=json
- limit=8
3. Save the raw JSON response exactly as returned to search-response.json in the current workspace
4. Do not transform fields and do not wrap with extra text
5. If successful, reply only DONE.
6. If command execution fails, reply only FAIL.
I did exactly that:
1. I read the contents of search-url.txt (http://127.0.0.1:34785/search/sn6IJOtP7uwGtxMPjP_EHmQt) and query.txt (weekly plan 5eb575f70aee checkbox)
2. I called the search endpoint with the correct parameters
3. I saved the raw JSON response to search-response.json
4. I replied only DONE as instructed
The task is complete.
</think>
</think>
DONE
[15:04:24] INFO Auto-reply: Assistant message had no tools. Asking for confirmation... complete.py:138
[15:04:24] [Loop control] <system>No tool call detected in last message. Did you mean to finish? If so, make sure you are comp...
User: <system>No tool call detected in last message. Did you mean to finish? If so, make sure you are completely done and then use the `complete` tool to end the session.</system>
Assistant: Thinking...
Assistant:
<think>
<think>
The user asked me to:
1. Read search-url.txt and query.txt from the current workspace
2. Call the search endpoint with query parameters:
- q=<urlencoded query from query.txt>
- format=json
- limit=8
3. Save the raw JSON response exactly as returned to search-response.json in the current workspace
4. Do not transform fields and do not wrap with extra text
5. If successful, reply only DONE.
6. If command execution fails, reply only FAIL.
I've already executed the curl command and saved the response to search-response.json. The command completed without errors (no output means success for curl with -s flag).
The task is complete. I should use the complete tool to end the session.
</think>
</think>
@complete(call_n5caduz7): {}
System: Task complete. Autonomous session finished.
[15:04:39] INFO Complete tool call detected, stopping session immediately complete.py:80
INFO Hook 'complete.complete' signaled session completion __init__.py:360
[15:04:39] Autonomous mode: Session completed via complete tool. Exiting.
Goodbye! (resume with: gptme --name gptme__local_aravhawk_qwen3_5_opus_4_6_text_9b__ollama__tool__web_fetch_json_raw-measured)
[gripprobe] process_finished_at=2026-05-01T15:04:40+00:00 exit_code=0
[gripprobe] process_started_at=2026-05-01T15:03:30+00:00 [gripprobe] process_finished_at=2026-05-01T15:04:40+00:00 exit_code=0
# Modelfile generated by "ollama show"
# To build a new Modelfile based on this, replace FROM with:
# FROM aravhawk/qwen3.5-opus-4.6-text:9b
FROM /usr/share/ollama/.ollama/models/blobs/sha256-d0ecd80b0e45b0d9e49c8cd1527b7f7d52d8d3bde2c569ab36aac59bb78f53f7
TEMPLATE {{ .Prompt }}
RENDERER qwen3.5
PARSER qwen3.5
PARAMETER num_ctx 32768
PARAMETER presence_penalty 1.5
PARAMETER stop <|im_end|>
PARAMETER stop <|im_start|>
PARAMETER temperature 1
PARAMETER top_k 20
PARAMETER top_p 0.95