Add files using upload-large-folder tool

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+eval-results/qa/EVAL_RESULTS.pdf filter=lfs diff=lfs merge=lfs -text

benchmark-scaling-audit/raw/20260804-234956-snowball-step105149-gsm8k-0shot-243e/0000000000000000-c65b047e.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b82a80af721faf70cf6f52b09825e76a193a32d78cce712e23792e2242bcc167
+size 585136

benchmark-scaling-audit/raw/20260804-234958-snowball-step105149-triviaqa-7177/0000000000000000-9d95cb48.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:835c78704833c3ffdec491759ad1308bafb2a3858ab204344fd158700918fc95
+size 6106656

benchmark-scaling-audit/raw/20260804-235006-snowball-step105149-nq-open-fd40/0000000000000000-b3238ac4.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3d84c61e5103c91129f184de177a4a1b50ed33f3734c3a87eec655c6231467f1
+size 301278

benchmark-scaling-audit/raw/20260804-235008-snowball-step105149-drop-93a9/0000000000000000-088a3718.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:596cf7f0b90bcf97c30611c8d052c7e5d8ca87fc689507476ce4088fda9cfb59
+size 1324713

benchmark-scaling-audit/raw/20260806-200908-snowball-sft-drop-0f11/0000000000000000-a8e1d82c.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b194b294c77fc12b1c17e6cb046e9b85db15c1b84e9c247a26e29acbe2505890
+size 3814260

benchmark-scaling-audit/raw/20260806-200908-snowball-sft-gsm8k-0shot-303f/0000000000000000-b5d5630f.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:32174bcfd22d8ba055348b223ea0a926449810e085bbd72c2732ae55f0f12b54
+size 1178391

benchmark-scaling-audit/raw/20260806-200908-snowball-sft-olympiadbench-489a/0000000000000000-92d71bb9.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c13669e9a80fda24761cae4eb99cc70c14c5dadb24e321a7579c8c0d08978c10
+size 141852

benchmark-scaling-audit/raw/20260807-120029-snowball-chat-olympiadbench-9398/0000000000000000-ee7d01b5.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c956af4d84bf7ca3d43798679ae23a2c8b82bec646c13918034348105cb3ee9
+size 110594

benchmark-scaling-audit/raw/20260807-120029-snowball-drop-a9ab/0000000000000000-b2173515.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f98e61cc3a591ea539d60f906809f74efebecdda7a32e1c8fc23a991113ff47f
+size 1311334

benchmark-scaling-audit/raw/20260807-120029-snowball-gsm8k-0shot-f50f/0000000000000000-40373ac4.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6ad30c05c171453e4d0de975f18321d8dde9c90f4c8a1db3899d01a7cbcdb419
+size 579897

benchmark-scaling-audit/raw/20260807-120029-snowball-nq-open-f611/0000000000000000-bac2804b.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be06574ab27950af3023949e4627f36119a12f3c6bd927d7e3f0a158ce1685bf
+size 301528

benchmark-scaling-audit/raw/20260807-120029-snowball-olympiadbench-1153/0000000000000000-07d2fbfd.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec43949e933af9874a989a47c162c2412da4bc001582a244127fb6798efcc529
+size 109771

benchmark-scaling-audit/raw/20260807-120029-snowball-triviaqa-67f1/0000000000000000-138796e0.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:588db6b1d3928f335a840fb3f94b2f5e6afcc0393004bef38e5e4a6412f0f186
+size 6114975

benchmark-scaling-audit/raw/20260808-192144-snowball-chat-drop-dd48/0000000000000000-bbc7867c.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dca82023eab22aa7a226d0432c1d5f91109fa6f6b75f204a2e82e329e75d2be2
+size 1481991

benchmark-scaling-audit/raw/20260808-192144-snowball-chat-gsm8k-0shot-29f1/0000000000000000-546e30d8.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a76b7d15bc444d2bb8822305ad5e557ed689b5119f0505718e795656c546365d
+size 823381

benchmark-scaling-audit/raw/20260813-185542-snowball-step105149-sft-s1-chat-drop-7de8/samples_drop_2026-08-13T21-00-05.376452.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:010e3284743c868affd966799ca0f0cd479ebafc92e1c035d67b2a8b447c1bba
+size 2939361

benchmark-scaling-audit/raw/20260813-185542-snowball-step105149-sft-s1-chat-gsm8k-0shot-19e5/samples_gsm8k_2026-08-13T21-04-37.413918.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1ccf924a88dd89c1b1d6d5d87c2e9a6620acce10afcba76910202cd2ba1726f6
+size 1312996

benchmark-scaling-audit/raw/20260813-185542-snowball-step105149-sft-s1-chat-nq-open-490a/samples_nq_open_2026-08-13T20-50-20.156979.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5254e8ee621986195ae02f3fd81b76f5c6d8dc21d5e64c9cfc25ceed968d6be5
+size 591048

benchmark-scaling-audit/raw/20260813-185542-snowball-step105149-sft-s1-chat-olympiadbench-d4df/samples_OlympiadBench_2026-08-13T19-20-20.846181.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c642f7ce78f2f57b7fb66513ae10bba8d0804c8ca6085712c641c7b7a4ce263b
+size 168726

benchmark-scaling-audit/raw/20260813-185542-snowball-step105149-sft-s1-chat-triviaqa-460a/samples_triviaqa_2026-08-13T20-46-46.062807.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:63336114531ffc747ac9cfb4e6d245024e01c0d5d595140725f34586365b94b9
+size 9391361

benchmark-scaling-audit/raw/20260813-235536-snowball-step105149-sft-s2-thinking-olympiadbench-03b8/samples_OlympiadBench_2026-08-14T00-47-19.235444.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ffb47ca6981f0a377882c42ce448758088e79ad3f1cbfb54873d535aa03b8d20
+size 266042

benchmark-scaling-audit/raw/20260814-135852-snowball-step105149-base-olympiadbench-e212/0000000000000000-905fb3e8.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:13bc3c8ed1e9ed00edb07c44961e2b395917debfebdd0db38b6f97adb6b0574b
+size 95246

benchmark-scaling-audit/raw/20260815-045136-snowball-step105149-sft-s3-nemotron-terminal-nq-open-58af/0000000000000000-7c85d6c2.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0f7bf06a9e7e09aa6dc8e9ef04c25dca2000fab47cddf8b3630673ef0f9bdeb
+size 3284356

benchmark-scaling-audit/raw/20260815-045136-snowball-step105149-sft-s3-nemotron-terminal-olympiadbench-d7bf/0000000000000000-610af0cb.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a6e4df8ff69fd4a7aad71e5b96d4ae8023dc62beeb2c245e303a6ea4f57f57f
+size 157314

benchmark-scaling-audit/raw/20260815-045136-snowball-step105149-sft-s3-nemotron-terminal-triviaqa-1ea2/0000000000000000-0eb4dcc5.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e4d3feac62cd45c319f35d196e441ed677b2cc1de3a00c6a2ca9bd37122cf77
+size 19536412

benchmark-scaling-audit/raw/20260815-142937-snowball-step105149-sft-s2-thinking-drop-cb68/0000000000000000-881c4342.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:841dd89747c612fb0d6893484aca82c4db7a3d74107ef2833fde6d7f48afbfe5
+size 6058330

benchmark-scaling-audit/raw/20260815-142937-snowball-step105149-sft-s2-thinking-gsm8k-0shot-48e3/0000000000000000-3afdfdf3.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ea339d684219754548a043f6ca55e7f2b867bb4e67c33fcb2a321a4f91e4eb9b
+size 1456583

benchmark-scaling-audit/raw/20260815-145138-snowball-2t-sft-repro-nemotron-terminal-step1888-olympiadbench-9738/0000000000000000-72debc4b.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2d0549c5c58cbb687ab01da8dd3c6f95cd508e581c53c8b3f1c53b478862f1f1
+size 162110

benchmark-scaling-audit/raw/20260815-145702-snowball-2t-chat-step257-nq-open-8e03/0000000000000000-1252a996.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a38e678ba49b296642b34a60f50ca38bda0ee0ff6103fa7036f06cff9de17c8
+size 795196

benchmark-scaling-audit/raw/20260815-145702-snowball-2t-chat-step257-triviaqa-c197/0000000000000000-9e83acf6.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d323b04e69cdcef62aeb7ef7c7488a676a5457bac034f0daf8419783a76e89b8
+size 6366108

benchmark-scaling-audit/raw/20260815-145702-snowball-2t-thinking-step630-nq-open-a6ad/0000000000000000-e9ab15c0.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a70ee8e28f74e1602addf9711bb0f29e7ddb78ec27e450d5773a497b6f68ebf
+size 4153102

benchmark-scaling-audit/raw/20260815-200045-snowball-2t-sft-repro-nemotron-terminal-step1888-2t-agentic-90cf/0000000000000000-dcd82ff7.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e46e80b9574f864fc9a6cb9929ab4187c4e9b9f4b6104ff9a42355ed4681cdc1
+size 3509928

benchmark-scaling-audit/raw/20260815-200045-snowball-2t-sft-repro-nemotron-terminal-step1888-2t-agentic-90cf/0000000000004876-b5e6ca97.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d6dcaa91a49c328b1d8c6a606b77d99e7d26ff4173c4c8919b29690a8657bbf
+size 3436255

benchmark-scaling-audit/raw/20260815-200045-snowball-2t-sft-repro-nemotron-terminal-step1888-2t-agentic-90cf/0000000000009536-fea59742.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:774f168a7dc3222becdcf2ccbb3067609c64561aa412f2f546506ab336fce55f
+size 1973854

benchmark-scaling-audit/raw/20260815-200045-snowball-2t-sft-repro-nemotron-terminal-step1888-2t-agentic-90cf/0000000000012174-d572e816.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ce6107b62d09654e881d1049bd245cec2a7fbbb9515f1503f6681a594528799
+size 1187800

benchmark-scaling-audit/raw/20260815-200045-snowball-step105149-sft-s2-thinking-stage2-126c/0000000000000000-d8fa74bb.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e6bb8a6f53be4ef795d9e831da0cdbc1aa8f9e8fbd89565771512d4a620c1cd1
+size 3839046

benchmark-scaling-audit/raw/20260815-200045-snowball-step105149-sft-s2-thinking-stage2-126c/0000000000000000-dfa4d65f.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:68c80f5de13072511ab77dce5ffb0ece2ebe16dbafdb7d18145eb553c4cd67e6
+size 215496

benchmark-scaling-audit/raw/20260815-200045-snowball-step105149-sft-s3-nemotron-terminal-stage3-8bcc/0000000000000000-a120398c.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c6678c687b48968a9ff8560b9349b3d3ced859ab24693379b6852e91078cc39
+size 8123860

benchmark-scaling-audit/raw/chat_drop_direct.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:010e3284743c868affd966799ca0f0cd479ebafc92e1c035d67b2a8b447c1bba
+size 2939361

benchmark-scaling-audit/raw/stage3_compact.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c6678c687b48968a9ff8560b9349b3d3ced859ab24693379b6852e91078cc39
+size 8123860

eval-configs/2t/sft-repro-thinking-step630-nemotron-terminal-step1888.yaml

@@ -0,0 +1,25 @@
+name: snowball-2t-sft-repro-nemotron-terminal-step1888
+location: s3://marin-us-east-02a/marin/grug/sft_repro_thinking_step630_nemotron_terminal_steps1888/2026.08.11.1/exports/step-1888/hf-bf16-vllm/
+tokenizer: marin-community/marin-tokenizer
+apply_chat_template: true
+resource_hint:
+  gpu:
+    H100: 8
+  memory: 512g
+serve:
+  tensor_parallel_size: 1
+  data_parallel_size: 8
+  max_model_len: 32768
+  max_num_batched_tokens: 8192
+  max_num_seqs: 8
+  vllm_extra_args:
+  - --enable-expert-parallel
+  - --model-loader-extra-config
+  - '{"distributed":true}'
+  - --gpu-memory-utilization
+  - '0.88'
+generation:
+  max_gen_toks: 8192
+  extra_gen_kwargs:
+    skip_special_tokens: "false"
+    repetition_penalty: "1.1"

eval-configs/2t/snowball-chat-step257.yaml

@@ -0,0 +1,28 @@
+name: snowball-2t-chat-step257
+location: s3://marin-us-east-02a/models/penfever--grug-67b-a2b-sft-s1-wildchat-step257/
+tokenizer: marin-community/marin-tokenizer
+apply_chat_template: true
+resource_hint:
+  gpu:
+    H100: 8
+  memory: 512g
+serve:
+  tensor_parallel_size: 1
+  data_parallel_size: 8
+  max_model_len: 32768
+  max_num_batched_tokens: 8192
+  max_num_seqs: 8
+  vllm_extra_args:
+  - --enable-expert-parallel
+  - --model-loader-extra-config
+  - '{"distributed":true}'
+  - --gpu-memory-utilization
+  - '0.88'
+generation:
+  max_gen_toks: 8192
+  extra_gen_kwargs:
+    skip_special_tokens: "false"
+    repetition_penalty: "1.1"
+agent:
+  agent_kwargs:
+    extra_body: '{"chat_template_kwargs":{"enable_thinking":false}}'

eval-configs/2t/snowball-thinking-step630.yaml

@@ -0,0 +1,28 @@
+name: snowball-2t-thinking-step630
+location: s3://marin-us-east-02a/marin/exports/grug/june-67b-a2b-sft-s2-thinking/step-630/hf-bf16-vllm/
+tokenizer: marin-community/marin-tokenizer
+apply_chat_template: true
+resource_hint:
+  gpu:
+    H100: 8
+  memory: 512g
+serve:
+  tensor_parallel_size: 1
+  data_parallel_size: 8
+  max_model_len: 32768
+  max_num_batched_tokens: 8192
+  max_num_seqs: 8
+  vllm_extra_args:
+  - --enable-expert-parallel
+  - --model-loader-extra-config
+  - '{"distributed":true}'
+  - --gpu-memory-utilization
+  - '0.88'
+generation:
+  max_gen_toks: 8192
+  extra_gen_kwargs:
+    skip_special_tokens: "false"
+    repetition_penalty: "1.1"
+agent:
+  agent_kwargs:
+    extra_body: '{"chat_template_kwargs":{"enable_thinking":false}}'

eval-configs/canonical-ot-tblite/ISSUE_7958_COMMENT.md

@@ -0,0 +1,21 @@
+🤖 #7958 did not preserve a complete OT-TBLite policy. The public gist below supplies a two-file launch tuple:
+
+- `ot-tblite.harbor.yaml` fixes the benchmark policy: `DCAgent/dev_set_v2@377118ff3031c934f5a647ae2c425eb74eef3b21`, Terminus-2, three attempts, 128 concurrent trials, Daytona automatic snapshots, linear raw trajectories, a 32,768-token input limit, a 16,384-token output limit, and the listener-resolved timeout settings.
+- `model.example.marin.yaml` shows the separate Marin `ModelConfig`. Replace its identity and serving geometry for the checkpoint under test; keep the Harbor policy fixed.
+
+https://gist.github.com/penfever/767c5d224377292df95e20035e0155f1
+
+The retry policy deliberately excludes only `AgentTimeoutError` and `ContextLengthExceededError`. Harbor retains and verifies those scoreable agent outcomes. Every other exception remains eligible for retry, including infrastructure types added after OpenThoughts-Agent's older config was written.
+
+The OpenThoughts-Agent `v2` preset currently resolves 32 concurrent trials. This tuple uses Marin's canonical 128-trial concurrency; the rest of the effective policy follows `unified_eval_listener.py`'s `v2.yaml` → `dcagent_eval_config_no_override.yaml` → runtime-overlay precedence.
+
+```bash
+env -u DAYTONA_API_KEY uv run python -m experiments.evaluation.cli launch \
+  --model-config model.example.marin.yaml \
+  --harbor-config ot-tblite.harbor.yaml \
+  --federated_cluster cw-rno2a \
+  --priority interactive \
+  --dry-run
+```
+
+Remove `--dry-run` after the resolved model, accelerator, target cluster, and records prefix are correct. Marin `main` at `90001550a095ea992c6a687e968600b598702904` accepts this tuple in preflight.

eval-configs/canonical-ot-tblite/model.example.marin.yaml

@@ -0,0 +1,18 @@
+# Example Marin ModelConfig. Replace model identity and serving geometry for the
+# checkpoint under test. Keep the Harbor file unchanged across model cohorts.
+name: qwen3-32b
+location: Qwen/Qwen3-32B
+apply_chat_template: true
+resource_hint:
+  gpu:
+    H100: 2
+serve:
+  tensor_parallel_size: 2
+  max_model_len: 32768
+  tool_call_parser: hermes
+  reasoning_parser: qwen3
+  vllm_extra_args:
+    - --enable-prefix-caching
+agent:
+  agent_kwargs:
+    extra_body: '{"chat_template_kwargs":{"enable_thinking":true}}'

eval-configs/canonical-ot-tblite/ot-tblite.harbor.yaml

@@ -0,0 +1,52 @@
+# Resolved equivalent of OpenThoughts-Agent's unified_eval_listener.py --preset v2.
+# Source precedence: eval/presets/v2.yaml, then
+# hpc/harbor_yaml/eval/configs/dcagent_eval_config_no_override.yaml, then the
+# listener/sbatch CLI overlays. Model name and endpoint are injected by Marin.
+# Marin canonical override: 128 concurrent trials; the listener v2 preset uses 32.
+n_attempts: 3
+timeout_multiplier: 1.0
+n_concurrent_trials: 128
+retry:
+  max_retries: 10
+  # These are scoreable agent outcomes. Every other exception is retryable.
+  exclude_exceptions:
+    - AgentTimeoutError
+    - ContextLengthExceededError
+  wait_multiplier: 1.0
+  min_wait_sec: 1.0
+  max_wait_sec: 60.0
+environment:
+  type: daytona
+  force_build: true
+  delete: true
+  kwargs:
+    auto_snapshot: true
+verifier:
+  max_timeout_sec: 14400
+agents:
+  - name: terminus-2
+    max_timeout_sec: 7200
+    kwargs:
+      trajectory_config:
+        raw_content: true
+        linear_history: true
+      record_terminal_session: false
+      enable_episode_logging: false
+      enable_pane_logging: false
+      collect_rollout_details: false
+      collect_engine_metrics: false
+      proactive_summarization_threshold: 2048
+      metrics_endpoint: https://replace-with-vllm-host/metrics
+      metrics_timeout_sec: 10
+      llm_kwargs:
+        timeout: 1800
+      model_info:
+        max_input_tokens: 32768
+        max_output_tokens: 16384
+        input_cost_per_token: 0.0
+        output_cost_per_token: 0.0
+      max_tokens: 16384
+datasets:
+  - name: hf://DCAgent/dev_set_v2
+    # main resolved to this revision when the canonical tuple was published.
+    ref: 377118ff3031c934f5a647ae2c425eb74eef3b21

eval-configs/consistent-drop-olympiad-20260816/2t-agentic.model.yaml

@@ -0,0 +1,25 @@
+name: snowball-2t-agentic-unified-20260816
+location: s3://marin-us-east-02a/marin/grug/sft_repro_thinking_step630_nemotron_terminal_steps1888/2026.08.11.1/exports/step-1888/hf-bf16-vllm/
+tokenizer: marin-community/marin-tokenizer
+apply_chat_template: true
+resource_hint:
+  gpu:
+    H100: 8
+  memory: 512g
+serve:
+  tensor_parallel_size: 1
+  data_parallel_size: 8
+  max_model_len: 32768
+  max_num_batched_tokens: 32768
+  max_num_seqs: 32
+  vllm_extra_args:
+  - --enable-expert-parallel
+  - --model-loader-extra-config
+  - '{"distributed":true}'
+  - --gpu-memory-utilization
+  - '0.88'
+generation:
+  max_gen_toks: 32768
+  extra_gen_kwargs:
+    skip_special_tokens: "false"
+    repetition_penalty: "1.1"

eval-configs/consistent-drop-olympiad-20260816/2t-base.model.yaml

@@ -0,0 +1,26 @@
+name: snowball-2t-base-unified-20260816
+location: s3://marin-us-east-02a/marin/exports/grug/june-67b-a2b/step-42150/hf-bf16-vllm/781bc3291c81ce28/
+tokenizer: marin-community/marin-tokenizer
+apply_chat_template: true
+resource_hint:
+  gpu:
+    H100: 8
+  memory: 512g
+serve:
+  tensor_parallel_size: 1
+  data_parallel_size: 8
+  max_model_len: 32768
+  max_num_batched_tokens: 32768
+  max_num_seqs: 32
+  chat_template: "{%- for message in messages -%}{{ message['content'] }}\n\n{%- endfor -%}"
+  vllm_extra_args:
+  - --enable-expert-parallel
+  - --model-loader-extra-config
+  - '{"distributed":true}'
+  - --gpu-memory-utilization
+  - '0.88'
+generation:
+  max_gen_toks: 32768
+  extra_gen_kwargs:
+    skip_special_tokens: "false"
+    repetition_penalty: "1.1"

eval-configs/consistent-drop-olympiad-20260816/2t-chat.model.yaml

@@ -0,0 +1,25 @@
+name: snowball-2t-chat-unified-20260816
+location: s3://marin-us-east-02a/models/penfever--grug-67b-a2b-sft-s1-wildchat-step257/
+tokenizer: marin-community/marin-tokenizer
+apply_chat_template: true
+resource_hint:
+  gpu:
+    H100: 8
+  memory: 512g
+serve:
+  tensor_parallel_size: 1
+  data_parallel_size: 8
+  max_model_len: 32768
+  max_num_batched_tokens: 32768
+  max_num_seqs: 32
+  vllm_extra_args:
+  - --enable-expert-parallel
+  - --model-loader-extra-config
+  - '{"distributed":true}'
+  - --gpu-memory-utilization
+  - '0.88'
+generation:
+  max_gen_toks: 32768
+  extra_gen_kwargs:
+    skip_special_tokens: "false"
+    repetition_penalty: "1.1"