Datasets:
File size: 2,366 Bytes
d6c2550 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 | {
"schema_version": 1,
"audit_date": "2026-10-21",
"source_dataset_revision": "1ca849c2128cf9ba9090a11514378c3a8277977a",
"scope": "published v2 evaluation",
"method": "Sample saved result.config from first and last row groups of first, middle and last trace files; one available result per model per selected row group. This is not an exhaustive attempt audit.",
"sampled_configs": 1,
"evaluation_settings": [
{
"model": "openai/coder-30b",
"agent": "terminus-2",
"context_tokens": 32768,
"intended_output_tokens_per_call": 8192,
"request_output_tokens": null,
"thinking": false,
"sampling": {
"temperature": 0.7,
"top_p": 0.8,
"top_k": 20,
"repetition_penalty": 1.05
},
"output_limit_status": "No explicit request cap in sampled saved configs; actual fallback cap unknown"
}
],
"evidence": [
{
"trace_file": "nl2bash_v2/traces/coder.parquet",
"row_group": 0,
"attempt_id": "0408302b-c0e9-4804-94da-07c1faba4873",
"model": "openai/coder-30b",
"agent_kwargs": {
"temperature": 0.7,
"extra_body": {
"top_p": 0.8,
"top_k": 20,
"repetition_penalty": 1.05,
"chat_template_kwargs": {
"enable_thinking": false
}
},
"max_tokens": 8192,
"model_info": {
"max_input_tokens": 32768,
"input_cost_per_token": 0,
"output_cost_per_token": 0,
"max_output_tokens": 8192
}
},
"resolved_timeouts": {
"agent": 600.0,
"verifier": 600.0,
"agent_setup": 360.0,
"environment_build": 600.0,
"trial_attempt": null,
"trial_cleanup_grace": 30.0,
"steps": []
},
"agent_override_timeout_sec": null
}
],
"actual_fallback_output_cap": null,
"findings": "Saved configurations use top-level max_tokens without llm_call_kwargs.max_tokens. The affected Terminus forwarding path does not send that top-level setting to the model API. No universal effective per-response token limit is established by this sample.",
"fixed_code": "config/models.py:serving_agent_kwargs forwards llm_call_kwargs.max_tokens; HTTP-request regression tests cover every current registry model; this does not change historical results."
}
|