Add muse-glimmer-30b/four-benchmarks-v2

.gitattributes

@@ -72,3 +72,4 @@ gemma-4-26b-a4b/four-benchmarks-v2/swe-mini/records.jsonl filter=lfs diff=lfs me
 gemma-4-26b-a4b/four-benchmarks-v2/tau3/records.jsonl filter=lfs diff=lfs merge=lfs -text
 qwen3.8-27b/four-benchmarks-v2/swe-mini/records.jsonl filter=lfs diff=lfs merge=lfs -text
 muse-glimmer-30b/four-benchmarks-v2/tau3/records.jsonl filter=lfs diff=lfs merge=lfs -text
+muse-glimmer-30b/four-benchmarks-v2/swe-mini/records.jsonl filter=lfs diff=lfs merge=lfs -text

muse-glimmer-30b/four-benchmarks-v2/README.md

@@ -1,17 +1,20 @@
 # Muse-Glimmer-30B on 125 tasks of each benchmark, corrected setup
 
-`meta-models/Muse-Glimmer-30B` on 125 tasks of each of two agent benchmarks, one run for each task. Batch job 3342576 on DeltaAI (node gh045, two GH200s), Oct 8, 2026, 56 min. The code is commit `aad53cb` of the team repo.
+`meta-models/Muse-Glimmer-30B` on 125 tasks of each of four agent benchmarks, one run for each task. Two batch jobs on DeltaAI did the work, with commit `aad53cb` of the team repo:
+
+- Job 3342576 (node gh045, two GH200s), Oct 8, 2026, 56 min: τ³-bench and AppWorld.
+- Job 3342577 (node gh033, one GH200), Oct 9, 2026, 2 h 7 min: SWE-bench and BrowseComp-Plus on one server.
 
 This is the second run of the model, after the corrections of the Oct 6 audit. One rule gives the 125 tasks of each benchmark: sort the task ids, then `random.Random(0).sample(ids, 125)`. The lists are in [`tasks-500/`](https://github.com/mithils3/distillation-stream/tree/main/lead/runs/qwen38-five/tasks-500) of the team repo.
 
-SWE-bench and BrowseComp-Plus are not in this folder at this time. Job 3342577 does them, and they come to this folder after that job. First run, no baseline. Training link unknown.
+First run, no baseline. Training link unknown.
 
 | Benchmark | Files | Tasks | Records | Result | Median tool calls |
 |---|---|---|---|---|---|
 | τ³-bench v1.0.1, `base` split, 58 banking and 67 telecom tasks | `tau3/` | 125 | 125 | 69 with reward 1.0 (11 banking, 58 telecom) | 10 |
 | AppWorld `42b5bcf`, 31 `test_normal` and 94 `test_challenge` tasks | `appworld.tar.gz.enc` | 125 | 125 | 90 pass each test | 21 |
-| SWE-bench Verified | not in this folder | 125 | 0 | comes after job 3342577 | |
-| BrowseComp-Plus | not in this folder | 125 | 0 | comes after job 3342577 | |
+| SWE-bench Verified, arm64 images, mini-SWE-agent 2.1.0 | `swe-mini/` | 125 | 125 | 83 resolved | 82 |
+| BrowseComp-Plus, BM25, harness `liuqi6777/search_agent` `51ceadc` | `bcp-agent.tar.gz.enc` | 125 | 125 | 72 stop with an answer, and no answer has a grade | 35 |
 
 ## Known problems
 
@@ -24,19 +27,25 @@ This run:
 - AppWorld: 98 of 2,755 steps failed. In 73 of them, an app refused a request (HTTP 401 in 46 steps, HTTP 422 in 27). In 19, the code used an argument name that the API does not have. The run `988af8e_2` stopped at the limit of 50 steps.
 - AppWorld: `failures.jsonl` in the archive has 16 lines from job 3330982. An error in our worker stopped 8 tasks two times each, before the first model call. Each of these tasks has a record from job 3342576.
 - The simulated customer of τ³-bench is Qwen3.8-27B, on the second GPU.
+- SWE-bench: 3 of 125 runs became larger than the window of 131,072 tokens, and they have no submission: `django__django-15280`, `sphinx-doc__sphinx-9602` and `sympy__sympy-19040`. The other 122 runs stop with a submission. The model caused this.
+- SWE-bench: 764 of 11,030 `bash` calls failed (6.9%). 296 of them have an empty output, and 60 are commands that the task image does not have. The model caused these. The longest run used 2,391 s in the agent.
+- BrowseComp-Plus: 49 of 125 runs became larger than the window. 3 runs stop at the reply limit of 16,384 tokens, and 1 run gives an answer with no answer tags. 0 of 3,953 calls failed.
+- BrowseComp-Plus: the first 8 lines of `records.jsonl` are from job 3330982 of Oct 8 (code commit `00336b1`). The `harness` field and the parser file are the same as in job 3342577.
 
 Each model of this run:
 
-- SWE-bench: a check with the reference patch of each task found 3 tasks to remove before you compare models. For `astropy__astropy-8707` and `matplotlib__matplotlib-20859`, the reference patch does not pass in our arm64 containers. `psf__requests-1724` passes with an empty patch. This applies when SWE-bench comes to this folder.
+- SWE-bench: a check with the reference patch of each task found 3 tasks to remove before you compare models. For `astropy__astropy-8707` and `matplotlib__matplotlib-20859`, the reference patch does not pass in our arm64 containers. `psf__requests-1724` passes with an empty patch. Without the three tasks, Muse Glimmer resolves 82 of 122.
 - τ³-bench: a paid judge is necessary for some checks, and we did not do them. 1 record has `nl_assertions_skipped` (`banking_knowledge/task_102`), and its reward is a maximum. After an empty reply or an error, tau2 starts the conversation again. The record holds only the last try.
-- Do not compare these scores with public leaderboards. These are different: the harness settings and the Qwen customer in τ³-bench. We also used no judge.
-- In the τ³-bench `steps`, `tokens_in` and `tokens_out` are the same on each tool call of one reply. Add them one time for each reply.
+- Do not compare these scores with public leaderboards. These are different: the harness settings, the Qwen customer in τ³-bench, the arm64 subset of SWE-bench and the BM25 search. We also used no judge.
+- In `steps`, `tokens_in` and `tokens_out` are the same on each tool call of one reply. Add them one time for each reply.
 
 ## What changed since `four-benchmarks-v1`
 
 - Each argument of a tool call gets the type that the tool schema gives. The server does this with `muse_typed_parser.py`, a parser plugin for vLLM, and the τ³-bench runner does the same check.
   - In `four-benchmarks-v1`, on the same 58 banking tasks, 389 of 1,627 calls failed. The `arguments` text of `call_discoverable_agent_tool` and `give_discoverable_user_tool` came to the tool as an object.
   - In this run, 0 of the 297 calls to these two tools failed. Each of the 571 id values is a string.
+  - BrowseComp-Plus: in the first run of the model, the document id of each of the 1,139 `visit` calls came to the tool as a number, and each call failed. In this run, each of the 100 document ids is a string, and no call failed.
+- SWE-bench containers: `localhost` has an address, and `/testbed` is a copy on the local disk. The image has no commits from after the base commit. Each reply has a limit of 32,768 tokens.
 - AppWorld: each prompt has the date of its task. In `four-benchmarks-v1`, the prompts had the date of the job. No step of this run names a date of October 2026.
 - No task has a time limit. The step limits and the reply limits stay.
 - τ³-bench has banking and telecom tasks only.
@@ -50,20 +59,25 @@ The audit and the corrections are in the team repo: [lead log](https://github.co
 - `tau3/failures.jsonl`: one line for each try of a task that stopped. Each of the 7 lines names a full window.
 - `tau3/raw.tar.gz`: all other files that the harness wrote for τ³-bench.
 - `appworld.tar.gz.enc`: the same files for AppWorld, encrypted.
+- `swe-mini/records.jsonl` and `swe-mini/raw.tar.gz`: the same files for SWE-bench. It has no `failures.jsonl`, because no task stopped.
+- `bcp-agent.tar.gz.enc`: the same files for BrowseComp-Plus, encrypted.
 - `serve.sh`: how the server started.
 
-## How to open the encrypted file
+## How to open the encrypted files
 
-The license of AppWorld asks that public copies of its data, and of data made from it, are encrypted. This keeps test data out of training sets. Keep the opened files out of public locations.
+The license of AppWorld asks that public copies of its data, and of data made from it, are encrypted. BrowseComp-Plus supplies its questions and its public runs encrypted. This keeps test data out of training sets. Keep the opened files out of public locations.
 
 ```bash
 openssl enc -d -aes-256-cbc -pbkdf2 -pass pass:distillation-stream -in appworld.tar.gz.enc | tar -xz
+openssl enc -d -aes-256-cbc -pbkdf2 -pass pass:distillation-stream -in bcp-agent.tar.gz.enc | tar -xz
 ```
 
-The result is `appworld/`, with `records.jsonl`, `failures.jsonl` and `raw/`.
+The results are `appworld/`, with `records.jsonl`, `failures.jsonl` and `raw/`, and `bcp-agent/`, with `records.jsonl` and `raw/`.
 
 ## Settings
 
 - **Server.** vLLM 0.30.0 with `--max-model-len 131072 --reasoning-parser muse_glimmer --tool-parser-plugin muse_typed_parser.py --tool-call-parser muse_glimmer_typed --gpu-memory-utilization 0.90 --max-num-seqs 64 --language-model-only --enable-auto-tool-choice --kv-offloading-size 32 --kv-offloading-backend native`. 131,072 tokens is the full window of the model. Thinking is on, at the strength that the template of the model gives. The τ³-bench runner sends no temperature, top-p or top-k, thus the values of the model apply.
 - **τ³-bench.** The `harness` field: `tau2-bench v1.0.1 (fc0055d) llm_agent + user_simulator, tau2's seed and no other sampling settings sent, arguments fitted to the tool schema`. The customer is Qwen3.8-27B. 24 tasks ran at the same time.
 - **AppWorld.** The `harness` field: `AppWorld simplified_react_code_agent (appworld 0.2.0.dev0, commit 42b5bcf3cd33), temperature 1`. A run has a limit of 50 steps, and a reply has a limit of 32,768 tokens. 32 tasks ran at the same time, on the same server as τ³-bench.
+- **SWE-bench.** The `harness` field: `mini-swe-agent 2.1.0, swebench.yaml with tool calls, step limit 250, command timeout 60 s, cost limit off, reply limit 32768 tokens, network host, python bashrc (bash -c), later commits removed, ApptainerInstanceEnvironment`. 12 tasks ran at the same time.
+- **BrowseComp-Plus.** The `harness` field: `liuqi6777/search_agent 51ceadc (visit_full, native tool calls), retriever BM25 (pyserini 1.2.0, Tevatron/browsecomp-plus-indexes bm25), k=5, snippet 512 cl100k tokens, visit returns the whole document, 128 turns, no time limit, max_tokens 16384, no sampling settings sent, earlier thinking sent back, call timeout 3600 s, prompt date 2026-10-04, arguments fitted to the tool schema`. No judge gives a grade. 16 tasks ran at the same time, on the same server as SWE-bench.

muse-glimmer-30b/four-benchmarks-v2/bcp-agent.tar.gz.enc

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ba178090a3985979a5895952c0f5609c7b75b588ff75e05fef3ff0c0f1bc2ac7
+size 31609328

muse-glimmer-30b/four-benchmarks-v2/swe-mini/raw.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c5e52c23440c58fed6186400c3309b52e9129289bc613932a819c16cbe0c6803
+size 10682654

muse-glimmer-30b/four-benchmarks-v2/swe-mini/records.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:74e2991ef1915c0b40cf8c8d10a9f5f93214ee8285ab8e06b8b66dca684d92fe
+size 21182565