Datasets:
|
Download README.md from Squadstack/conversational-llm-benchmark: direct link, hf CLI and curl.
- Browser
- Download file 36.4 kB
-
https://huggingface.co/datasets/Squadstack/conversational-llm-benchmark/resolve/main/README.md
- Command line
-
hf download hf://datasets/Squadstack/conversational-llm-benchmark/README.md
-
curl -L -H "Authorization: Bearer $HF_TOKEN" -o README.md https://huggingface.co/datasets/Squadstack/conversational-llm-benchmark/resolve/main/README.md
36.4 kB
| pretty_name: "SquadStack Conversational LLM Benchmark for Voice Agents" | |
| license: other | |
| license_name: llm-benchmark-evaluation-licence-1.0 | |
| license_link: LICENSE.md | |
| language: | |
| - hi | |
| - en | |
| - ta | |
| - te | |
| - kn | |
| language_details: "hi-IN, en-IN, ta-IN, te-IN, kn-IN" | |
| language_creators: | |
| - found | |
| annotations_creators: | |
| - machine-generated | |
| multilinguality: | |
| - multilingual | |
| task_categories: | |
| - text-generation | |
| size_categories: | |
| - 1K<n<10K | |
| source_datasets: | |
| - original | |
| tags: | |
| - benchmark | |
| - evaluation-only | |
| - llm-evaluation | |
| - voice-agents | |
| - conversational-ai | |
| - turn-prediction | |
| - tool-calling | |
| - llm-judge | |
| - pass-rate | |
| - hinglish | |
| - code-switching | |
| - code-mixing | |
| - indian-languages | |
| - hindi | |
| - telesales | |
| - lead-qualification | |
| - csat-survey | |
| - latency | |
| configs: | |
| - config_name: test | |
| data_files: | |
| - split: test | |
| path: data/test/test-* | |
| - config_name: results | |
| data_files: | |
| - split: test | |
| path: data/results/test-* | |
| extra_gated_heading: "Request access to this benchmark" | |
| extra_gated_description: >- | |
| This dataset contains text reconstructed from real telephone conversations with Indian consumers. | |
| Access requests are reviewed manually; expect 3-5 working days. | |
| extra_gated_prompt: >- | |
| By requesting access you agree to the Terms of Use | |
| (TERMS_OF_USE.md). In particular you agree that you will not attempt to | |
| identify any caller or recover any redacted text, will not redistribute | |
| the data, and will not use the data to train, fine-tune or distil any model. | |
| extra_gated_fields: | |
| Name: text | |
| Affiliation: text | |
| Country: country | |
| Intended use: | |
| type: select | |
| options: | |
| - Research | |
| - Commercial evaluation of an LLM or voice-agent system | |
| - Education | |
| - label: Other | |
| value: other | |
| I agree not to attempt to identify any caller or recover any redacted text: checkbox | |
| I agree not to redistribute the data: checkbox | |
| I agree not to train, fine-tune or distil any model on this data: checkbox | |
| extra_gated_button_content: "Agree and request access" | |
| <picture> | |
| <source media="(prefers-color-scheme: dark)" srcset="https://huggingface.co/datasets/Squadstack/conversational-llm-benchmark/resolve/main/assets/01-at-a-glance-dark.png"> | |
| <img alt="At a glance: 471 calls across 6 campaign domains, 9,982 evaluation turns, 6 reply languages, 11 systems scored, 7 rubric parameters judged Pass or Fail on every turn, 9,962 turns scored." src="https://huggingface.co/datasets/Squadstack/conversational-llm-benchmark/resolve/main/assets/01-at-a-glance-light.png"> | |
| </picture> | |
| # SquadStack Conversational LLM Benchmark for Voice Agents | |
| **Version 1.0.0 · maintained by SquadStack** | |
| [Methodology](docs/METHODOLOGY.md) · [Curation](docs/CURATION.md) · [Datasheet](docs/DATASHEET.md) · [Leaderboard](docs/LEADERBOARD.md) · [Latency](docs/LATENCY.md) · [Prompts](docs/PROMPTS.md) · [Submit a system](docs/SUBMISSION.md) · [Licence](LICENSE.md) · [Terms of use](TERMS_OF_USE.md) | |
| ## Key takeaways | |
| **What this is.** 9,982 turns from 471 real Indian outbound voice-agent calls across six campaign domains. Each turn carries the customer's and the agent's messages up to that point. The systems also saw the runtime's ASR-confidence, silence and language messages and the tool turns, which are not shipped. The system under test writes the next agent turn, and a pinned LLM judge (Claude Haiku 5.5) labels it `Pass` or `Fail` on seven parameters. The question it answers: *which LLM should run inside an Indian voice agent, judged on whether its next turn follows the campaign, and how fast it starts speaking?* | |
| **What the run shows:** | |
| - **The in-house system leads, and no other system is tied with it.** Maya V1 passes 96.7% of verdicts (95% interval 96.4 to 97.0). The next two, Gemma 4 31B at 92.5% and DeepSeek V4.1 Flash at 92.2%, are close to each other and 4 to 5 points behind. Maya V1 was tuned on these campaigns and answers from a history its own production version wrote, which is why it is marked ★ and why the training-overlap check in [What would make these numbers wrong](#what-would-make-these-numbers-wrong) matters. | |
| - **The aggregate hides where a system fails.** GPT-6 Luna matches the open-weight models on the four mandates (93.1 mean) but passes `Tools` on only 70.0% of turns: it ends calls that should continue. Gemma 4 26B, GPT-4.1, Gnani Evon V3.3 30B A3B, Sarvam 105B, Nemotron 3.5 Lightning and Sarvam 30B fail `Goal` and `Context` most, which is drifting off the campaign and losing the place in the flow, while passing `Tools` at 86 to 95%. The two Gemma 4 sizes are 11 points apart overall, and about 26 points apart on `Goal`. | |
| - **The optimized prompt helps every system.** Overall pass-rate is 0.3 to 2.4 points higher under the Optimized prompt variant than under Normal, for all eleven systems ([PROMPTS.md](docs/PROMPTS.md)). | |
| - **The fastest endpoints are not the best.** Sarvam 105B and Sarvam 30B start speaking in about 270 ms, as fast as Maya V1, but pass 72.4% and 65.7% of verdicts. Sarvam 105B sits on the Pareto frontier only because nothing is faster; Sarvam 30B is slower and less accurate than it, so it does not. Latency compares hosting paths as much as models. | |
| - **Nearly half the turns follow uncertain speech recognition.** 45% of turns come straight after a low-confidence or empty-ASR message, the condition a voice agent meets and a text benchmark never does. | |
| ### Overall pass-rate | |
| <picture> | |
| <source media="(prefers-color-scheme: dark)" srcset="https://huggingface.co/datasets/Squadstack/conversational-llm-benchmark/resolve/main/assets/08-leaderboard-overall-dark.png"> | |
| <img alt="Overall pass-rate for 11 systems on 9,962 turns, with 95% intervals over calls. Maya V1, SquadStack in-house, leads at 96.7%; no other system is tied with it." src="https://huggingface.co/datasets/Squadstack/conversational-llm-benchmark/resolve/main/assets/08-leaderboard-overall-light.png"> | |
| </picture> | |
| *Overall pass-rate, higher is better: `Pass` verdicts as a share of all verdicts over the seven parameters. Solid bar = overall pass-rate with its 95% interval; pale bar = mean of the four mandate pass-rates; ★ marks SquadStack's in-house system. Full tables in [LEADERBOARD.md](docs/LEADERBOARD.md).* | |
| ### Latency against quality | |
| <picture> | |
| <source media="(prefers-color-scheme: dark)" srcset="https://huggingface.co/datasets/Squadstack/conversational-llm-benchmark/resolve/main/assets/09-latency-quality-pareto-p50-dark.png"> | |
| <img alt="P50 time to first token against overall pass-rate per system, with the Pareto frontier." width="960" src="https://huggingface.co/datasets/Squadstack/conversational-llm-benchmark/resolve/main/assets/09-latency-quality-pareto-p50-light.png"> | |
| </picture> | |
| *Overall pass-rate against P50 time to first token; the P90 and P95 views are in `assets/`. Maya V1 ran on a dedicated SquadStack endpoint; GPT-6 Luna and GPT-4.1 on the OpenAI API; Gemma, DeepSeek, Qwen and Nemotron through OpenRouter providers; the Gnani and Sarvam models on Hugging Face Inference Endpoints running vLLM. Latency therefore compares hosting paths as much as models. The clock and setup are in [LATENCY.md](docs/LATENCY.md).* | |
| **It is an evaluation set. It is not for training, and training on it is a licence violation.** | |
| --- | |
| ## What this measures — and what it does not | |
| | It measures | It does not measure | | |
| |---|---| | |
| | The next agent turn in a live outbound call, under a campaign prompt the agent must follow | Open-ended chat, creative writing, code | | |
| | Six campaign intents: loan sales, courier qualification, business-listing capture, admissions counselling, home-services booking, post-trip survey | Any campaign type not represented here | | |
| | Replies in Hindi, Hinglish, English, Tamil, Telugu and Kannada | The other scheduled Indian languages | | |
| | Behaviour after uncertain ASR, silences and language switches, as the runtime reported them | Speech recognition itself (see the [Conversational Streaming ASR Benchmark](https://huggingface.co/datasets/Squadstack/conversational-streaming-asr-benchmark)) | | |
| | Whether a turn follows the campaign: on goal, within the rules and facts, in the right place in the flow, based on what the customer said, the right scripted response, fit for speech, the right tool decision | String overlap with the production agent's reply | | |
| | Time to first token, on the setup in [LATENCY.md](docs/LATENCY.md) | Throughput, cost at scale, multi-turn drift (each turn is judged on the production history) | | |
| | | Adversarial behaviour: there is no adversarial set in this release | | |
| If your question is "which LLM should run inside my Indian voice agent", this benchmark is built for you. If your question is "which LLM is best", it is not: use it alongside general benchmarks such as [MMLU](https://arxiv.org/abs/2009.03300) and [MT-Bench](https://huggingface.co/spaces/lmsys/mt-bench) and Indic generation sets such as [IndicGenBench](https://huggingface.co/datasets/google/IndicGenBench_flores_in), not instead of them. | |
| --- | |
| ## At a glance | |
| | | | | |
| |---|---| | |
| | **Calls** | 471 real calls from six SquadStack campaigns | | |
| | **Turns** | 9,982 rows: 7,091 distinct turns, 2,891 of them under both prompt variants · median 14 distinct turns per call (P10 6, P90 25) | | |
| | **History** | median 18 customer and agent messages before the turn (P10 4, P90 43, max 388); median 26 (max 523) with the runtime messages and tool turns the systems saw, which are not shipped | | |
| | **Domains** | BFSI 189 calls · Marketplace 93 · Education 48 · Services 48 · Logistics 48 · Travel 45; one client campaign each | | |
| | **Target languages** | Hinglish 67% · Hindi 16% · English 8% · Tamil 6% · Kannada 1% · Telugu 1% (script of the production reply) | | |
| | **Turn conditions** | 45% after uncertain ASR · 444 silent-caller turns · 83 language switches | | |
| | **Split** | `test` only. There is no train split, by design. | | |
| | **Systems** | 11: 1 in-house, 2 commercial API, 8 open weights | | |
| | **Judge** | Claude Haiku 5.5, temperature 0, reasoning off, one rubric, `Pass` or `Fail` on every parameter | | |
| | **Headline metric** | Overall pass-rate over the seven parameters, with per-parameter pass-rates, slices and TTFT alongside | | |
| --- | |
| ## Dataset structure | |
| Two configs, joined on `sample_id` and `prompt_variant`. Loading the dataset without a config name gives you `test`. | |
| ```python | |
| from datasets import load_dataset | |
| bench = load_dataset("Squadstack/conversational-llm-benchmark", "test")["test"] | |
| res = load_dataset("Squadstack/conversational-llm-benchmark", "results")["test"] | |
| ``` | |
| | Config | Rows | Columns | What it is | | |
| |---|---:|---:|---| | |
| | `test` | 9,982 | 6 | the benchmark: one row per turn and prompt variant, with the call history up to the turn | | |
| | `results` | 9,962 | 105 | one row per scored turn: for each of the 11 systems, its reply, the judge's verdict on the seven parameters and its time to first token | | |
| The split is deliberate. `test` is the thing being released and does not change when a system is added or re-run; `results` is a snapshot of eleven systems and will. A new baseline never touches the benchmark itself. | |
| > **What is not in the files.** The campaign system prompts (643, one per call and variant) are supplied to gate-approved evaluators on request; a turn's prompt is the one for its `call_sid` and `prompt_variant` ([PROMPTS.md](docs/PROMPTS.md)). The production agent's reply to the turn is not shipped. The judge's one-sentence reason for each verdict is stored and available to approved evaluators. | |
| ### Recomputing our numbers | |
| Every accuracy figure on this card is a count over the verdicts in `results`: | |
| ```python | |
| P = ["Goal", "Safety", "Context", "Comprehension", "Response", "Delivery", "Tools"] | |
| s = "maya_v1" | |
| verdicts = [r[f"{s}__{p}"] for r in res for p in P if r[f"{s}__{p}"] is not None] | |
| overall_pass_rate = 100 * verdicts.count("Pass") / len(verdicts) | |
| ``` | |
| Our scoring pipeline produces the whole leaderboard, the intervals and [LEADERBOARD.md](docs/LEADERBOARD.md) from the same file. If your numbers disagree with ours, we treat it as our bug and log it as an erratum. | |
| ### Fields | |
| <details> | |
| <summary><b><code>test</code>: 6 fields, one row per turn and prompt variant</b></summary> | |
| | Field | Type | Description | | |
| |---|---|---| | |
| | `sample_id` | string | `sbllm-<first 8 characters of call_sid>-<turn>`. Pair with `prompt_variant`: 2,891 turns appear under both. Cite this, never a row index. | | |
| | `call_sid` | string | Call identifier. Aggregate and resample by call, never by turn. | | |
| | `prompt_variant` | string | `normal` or `optimized`: which campaign system prompt the turn was run under ([PROMPTS.md](docs/PROMPTS.md)) | | |
| | `code` | string | Campaign domain: `bfsi`, `marketplace`, `education`, `services`, `logistics`, `travel` | | |
| | `turn_index` | int32 | 1-based position of the turn in the call | | |
| | `previous_content` | list of messages | The customer's and the agent's messages before the turn, in OpenAI chat format: `role` (`user` for the customer, `assistant` for the agent) and `content`. The systems saw this history with the runtime's `system` messages (ASR confidence, empty ASR, silence, hold and language instructions) and the tool calls and results in place; those are not shipped. A history that ends on an `assistant` message (728 rows) is a turn where the agent speaks again without a new customer message, most often after a silence. | | |
| </details> | |
| <details> | |
| <summary><b><code>results</code>: one row per scored turn, 9 columns per system</b></summary> | |
| | Field | Type | Description | | |
| |---|---|---| | |
| | `sample_id`, `prompt_variant`, `call_sid`, `code`, `turn_index` | | Join keys to `test` | | |
| | `target_language` | string | Script of the production reply: `Hinglish`, `Hindi`, `English`, `Tamil`, `Telugu`, `Kannada`, or `Other` for a tool-only reply with no text | | |
| | `<system>__<Parameter>` | string | `Pass` or `Fail` for each of the seven parameters; missing when the judge returned no verdict for that system on that turn | | |
| | `<system>__ttft_ms` | int32 | Time to first content token for that turn; missing when it was not recorded | | |
| | `<system>__response` | string | The system's reply as the judge saw it, redacted like the histories (see [Redaction](#redaction)). A tool call is written `[tool_call] name({...})`. 241 replies that reproduced or quoted the system prompt read `[reply withheld: contained system-prompt text]`; their verdicts and time to first token are kept. | | |
| The 11 system prefixes, in column order: `deepseek_v4_1_flash`, `gemma_4_26b`, `gemma_4_31b`, `gnani_gnani_evon_v3_3_30b_a3b_fp8`, `gpt_4_1`, `gpt_6_luna`, `maya_v1`, `nemotron_3_5_lightning`, `qwen3_8_27b`, `sarvamai_sarvam_105b_fp8`, `sarvamai_sarvam_30b_fp8`. Display names, vendors, categories, model ids and hosting are in [`data/results/systems.json`](data/results/systems.json). Adding a system adds nine columns and changes nothing else. | |
| </details> | |
| <details> | |
| <summary><b>Other files in <code>data/</code></b></summary> | |
| | File | Holds | | |
| |---|---| | |
| | `data/results/leaderboard.json` | one object per system: rates, intervals, per-parameter rates, TTFT percentiles, rank, tie and Pareto flags | | |
| | `data/results/slice_rates.json` | overall pass-rate per system by domain and by target language, with turn and call counts and the 40-call floor flag | | |
| | `data/results/rubric.json` | the judge, the seven parameters, the rules and the aggregation, as run | | |
| | `data/results/systems.json` | display name, vendor, category, model id, hosting, provider and quantization per system | | |
| | `data/stats/corpus_stats.json` | composition of `test`, the source of [CURATION.md](docs/CURATION.md) | | |
| | `data/prompts/token_counts.json` | average system-prompt tokens per domain and variant, the source of [PROMPTS.md](docs/PROMPTS.md) | | |
| </details> | |
| ### Redaction | |
| Personal information and every company name are removed from the histories and from the system replies, and replaced by a placeholder: `[NAME]` for any person's name (customer, family member, third party), `[PHONE]`, `[EMAIL]`, `[PINCODE]`, `[ADDRESS]` for anything below city level, `[ID]` for account, application, booking and similar numbers, OTPs and promo codes, `[NUMBER]` for a business's age or outlet count in business-listing calls, and `[COMPANY]` for every client, product, lender, bus operator, courier, competitor, app, platform and business name, including inside tool and stage names (`fetch_<brand>_slots` becomes `fetch_company_slots`). Cities and states are kept. A check for facts that identify only in combination then replaced, in business-listing calls, the number of years the caller's business has run and one outlet count by `[NUMBER]`, and the small towns where two callers placed their own business by `[ADDRESS]`; in five travel calls it replaced journey times and one journey date by `[DATE]`. The runtime's system messages and the tool calls and results are removed from the histories, which keep only the customer's and the agent's text. No call timestamps are shipped; where a system reply copied the call's date and time from its prompt, it is replaced by `[DATE]`. Some replies also contain bracketed or braced tokens the systems wrote themselves (`[CITY]`, `[PRODUCT]`, `[COUNTERPARTY]`, `{lead_name}`); those are template echoes, not redaction. A system reply that reproduced or quoted the campaign or runtime prompt (241 of 109,500, mostly long multi-section outputs) is withheld: its text is replaced by `[reply withheld: contained system-prompt text]`, and its verdicts and time to first token are kept, so every score still recounts. | |
| The systems and the judge saw the **unredacted** prompts and histories, and the judge saw the unredacted replies, so the published scores are scores on the unredacted text. Every verdict is shipped, so counting them reproduces every number exactly; a re-run on the shipped files is a re-run on different input: redacted, and without the runtime messages and tool turns. The redaction method and counts are in [DATASHEET.md](docs/DATASHEET.md). It combines rule-based passes with a full read-through of every call, and it is not perfect: if you find something that should have been redacted, do not use it, and report the `sample_id` ([TERMS_OF_USE.md](TERMS_OF_USE.md)). | |
| --- | |
| ## How the 471 were chosen | |
| 471 real calls from six SquadStack campaigns, one per domain (BFSI 189, Marketplace 93, Education 48, Services 48, Logistics 48, Travel 45). There are no adversarial calls in this release. | |
| The calls are a quota sample per campaign, not a random one; the composition that can be checked from the files is in [`docs/CURATION.md`](docs/CURATION.md). | |
| The histories were produced by SquadStack's production agent. Overlap between these calls and customers and the data used to train SquadStack's in-house model was checked at call and customer level before release (see [What would make these numbers wrong](#what-would-make-these-numbers-wrong)). | |
| --- | |
| ## Diversity | |
| <picture> | |
| <source media="(prefers-color-scheme: dark)" srcset="https://huggingface.co/datasets/Squadstack/conversational-llm-benchmark/resolve/main/assets/03-diversity-dark.png"> | |
| <img alt="Panels: turns by domain and target language, calls and turns per domain, history depth, distinct turns per call, message roles in the histories, and the share of turns after uncertain ASR, silent callers, language switches and under both prompt variants." src="https://huggingface.co/datasets/Squadstack/conversational-llm-benchmark/resolve/main/assets/03-diversity-light.png"> | |
| </picture> | |
| ### Domain × target language | |
| Tamil, Telugu and Kannada replies sit almost entirely in Travel (556 of 846 turns), with most of the rest in BFSI. **A language slice is therefore mostly a campaign slice.** Read the cells, not the margins. The language of a turn is taken from the script of the production reply (Devanagari only = Hindi, Latin only = English, both = Hinglish), which is a proxy, not a language-ID model. | |
| ### Campaign | |
| Six domains, one client campaign each, so a domain slice cannot separate the vertical from its prompt. BFSI holds 41% of the turns; Travel is the only domain with all six languages in volume. | |
| ### What the agent is facing at the turn | |
| 45% of turns come straight after the runtime reported low-confidence or empty ASR, 444 after a silent caller and 83 after a language switch. The shipped histories hold 125k customer and 101k agent messages. The systems also saw 102k runtime messages and 3.9k tool results, which are not shipped, so these three conditions cannot be recomputed from `test`. | |
| ### Vocabulary | |
| Customer messages hold 63,539 tokens and 4,845 distinct words; 47% of the distinct words occur once, and MATTR-50 is 0.700. Composition tables per domain, and the method, are in [`docs/CURATION.md`](docs/CURATION.md). | |
| --- | |
| ## Why we judge turns, not strings | |
| <picture> | |
| <source media="(prefers-color-scheme: dark)" srcset="https://huggingface.co/datasets/Squadstack/conversational-llm-benchmark/resolve/main/assets/17-two-replies-dark.png"> | |
| <img alt="A loan-campaign customer asks what the EMI on two lakh over 24 months would be. Reply A says it will calculate and calls the EMI tool: all seven parameters pass. Reply B quotes an EMI of 8,500 rupees at a fixed 10% rate that nothing supplied: Rules & Facts fails." src="https://huggingface.co/datasets/Squadstack/conversational-llm-benchmark/resolve/main/assets/17-two-replies-light.png"> | |
| </picture> | |
| | Customer asked | Reply | Verdict | What happens next | | |
| |---|---|---|---| | |
| | EMI on ₹2 lakh over 24 months | "I'll calculate your EMI, one second" + `calculate_emi` | 7 of 7 Pass | The quoted number comes from the tool. | | |
| | EMI on ₹2 lakh over 24 months | "The EMI will be just ₹8,500, interest 10% fixed" | Rules & Facts Fail | The customer hears an offer nobody made. | | |
| Both replies are fluent Hinglish, on topic and the length of a voice turn, and a string-overlap metric against the production reply would score them alike. A voice agent's reply is spoken to a customer who acts on it, so the rubric asks whether the reply follows the campaign, and the Claims rule fails any unsupported commitment about money, eligibility or outcomes. *(Illustrative example written for this card, not a real call.)* | |
| --- | |
| ## Evaluation protocol | |
| <picture> | |
| <source media="(prefers-color-scheme: dark)" srcset="https://huggingface.co/datasets/Squadstack/conversational-llm-benchmark/resolve/main/assets/12-evaluation-pipeline-dark.png"> | |
| <img alt="Four steps: freeze the prompt and history, generate one reply with pinned decoding, judge seven parameters Pass or Fail with root-cause attribution, aggregate to the overall pass-rate with intervals over calls." src="https://huggingface.co/datasets/Squadstack/conversational-llm-benchmark/resolve/main/assets/12-evaluation-pipeline-light.png"> | |
| </picture> | |
| Four things about this protocol are load-bearing, and all are specified in [`docs/METHODOLOGY.md`](docs/METHODOLOGY.md). | |
| **The same input and the same judge for every system.** Every system sees the same campaign prompt and history for every turn, with declared decoding settings (temperature 0.7, 512 output tokens, reasoning off), and one pinned judge, Claude Haiku 5.5 at temperature 0, scores every reply with the same rubric. The judge's family has no system on the board. | |
| **One failure, one owner.** Each rule in a campaign prompt belongs to exactly one parameter. When a misheard customer leads to an off-goal reply, the failure is scored at `Comprehension` and `Goal` stays `Pass`. A violation is never counted twice. | |
| **No partial credit.** Every parameter is `Pass` or `Fail`. A paraphrase that keeps every required fact passes; a reply that drops a fact, adds an unsupported commitment or breaks a rule fails. A conditional parameter whose trigger did not fire is `Pass`. | |
| **Each turn is judged on its own,** against the production history, with no sight of the system's other turns or of the production reply. Every verdict and every reply ships in `results`; the judge's one-sentence reason per verdict is stored and available to approved evaluators. | |
| ### The seven parameters | |
| | Column | Rubric name | Type | What fails it | | |
| |---|---|---|---| | |
| | Goal | Goal Alignment | Mandate | The reply does not move the call toward the campaign goal, or wanders off scope | | |
| | Safety (shown as Rules & Facts) | Instruction, Safety & Verbatim Residual | Mandate | A scripted line loses a required fact, an unsupported commitment or false claim is made, privacy is breached, or a hard rule no other parameter owns is broken | | |
| | Context | Flow, Journey & Context | Mandate | Wrong stage, a restarted flow, a forgotten or re-asked value, misapplied reference data | | |
| | Comprehension | Perception & Understanding | Mandate | The reply acts on something the customer did not say: misread intent, mishandled low-confidence words, silence, digits or times | | |
| | Response | Response Selection | Conditional: a scenario, FAQ, objection, nudge or pitch was triggered | The wrong scripted response was chosen | | |
| | Delivery | Delivery Form | Conditional: the reply has spoken content | Wrong language or script, numbers not spoken as speech, too long, more than one action in a turn | | |
| | Tools | Tool Use | Every turn | A tool was needed and not called, the wrong tool, bad arguments, or a hang-up while the call should continue | | |
| ### The metrics | |
| | Metric | Definition | | |
| |---|---| | |
| | **Overall pass-rate** *(headline)* | `Pass` verdicts ÷ all verdicts, over the seven parameters and every judged turn | | |
| | **Mean of four mandates** | average of the four mandate pass-rates | | |
| | **Per-parameter pass-rate** | `Pass` turns ÷ judged turns, per parameter | | |
| | **Slices** | overall pass-rate by campaign domain and by target language | | |
| | **TTFT** | time from request to the first content token, P50, P90 and P95 over turns, on the setup in [LATENCY.md](docs/LATENCY.md) | | |
| Intervals are 95% bootstrap intervals over calls, never over turns, and differences between systems are paired on the same calls. Systems whose difference interval contains zero are reported as tied. A slice with fewer than 40 calls is descriptive only. | |
| --- | |
| ## Results | |
| One list, ordered by overall pass-rate. SquadStack's in-house system is marked ★ and ranked with the others: grouping an in-house model tuned on these campaigns separately from general-purpose endpoints is the usual way vendor benchmarks mislead. | |
| | # | System | Vendor | Category | Overall pass-rate | 95% CI | Mean of 4 mandates | Tied with #1 | Turns judged | TTFT P50 (ms) | | |
| |---:|---|---|---|---:|---|---:|:-:|---:|---:| | |
| | 1 | **Maya V1 ★** | SquadStack | In-house | **96.71** | 96.4–97.0 | 97.49 | — | 9,962 | 278 | | |
| | 2 | Gemma 4 31B | Google | Open weights | **92.45** | 92.1–92.8 | 93.47 | no | 9,962 | 652 | | |
| | 3 | DeepSeek V4.1 Flash | DeepSeek | Open weights | **92.18** | 91.8–92.5 | 93.27 | no | 9,959 | 604 | | |
| | 4 | GPT-6 Luna | OpenAI | Commercial API | **89.04** | 88.5–89.5 | 93.11 | no | 9,962 | 1,634 | | |
| | 5 | Qwen3.8 27B | Alibaba | Open weights | **83.77** | 83.1–84.4 | 88.15 | no | 9,962 | 1,061 | | |
| | 6 | Gemma 4 26B | Google | Open weights | **81.65** | 80.8–82.5 | 79.10 | no | 9,760 | 338 | | |
| | 7 | GPT-4.1 | OpenAI | Commercial API | **77.49** | 76.4–78.5 | 77.14 | no | 9,723 | 2,121 | | |
| | 8 | Gnani Evon V3.3 30B A3B | Gnani.ai | Open weights | **76.01** | 75.1–76.9 | 72.90 | no | 9,727 | 438 | | |
| | 9 | Sarvam 105B | Sarvam AI | Open weights | **72.41** | 71.5–73.3 | 70.03 | no | 9,523 | 270 | | |
| | 10 | Nemotron 3.5 Lightning | NVIDIA | Open weights | **71.28** | 70.3–72.3 | 69.42 | no | 9,614 | 690 | | |
| | 11 | Sarvam 30B | Sarvam AI | Open weights | **65.69** | 64.7–66.7 | 62.52 | no | 9,667 | 277 | | |
| *Generated by the scoring pipeline; 2,000 bootstrap replicates over calls, seed 20261008. GPT-4.1 has no verdict on 239 turns, Gnani Evon V3.3 30B A3B on 235, Sarvam 105B on 439, Nemotron 3.5 Lightning on 348, Sarvam 30B on 295, Gemma 4 26B on 202 and DeepSeek V4.1 Flash on 3 (the judge returned no readable verdict after five passes); those turns are not in that system's denominator. 20 turns of the 9,982 follow a call-ending tool and were not run.* | |
| ### What the aggregate hides | |
| <picture> | |
| <source media="(prefers-color-scheme: dark)" srcset="https://huggingface.co/datasets/Squadstack/conversational-llm-benchmark/resolve/main/assets/10-pass-by-parameter-dark.png"> | |
| <img alt="Heatmap of pass-rates for 11 systems on the overall rate and the seven parameters." src="https://huggingface.co/datasets/Squadstack/conversational-llm-benchmark/resolve/main/assets/10-pass-by-parameter-light.png"> | |
| </picture> | |
| Two systems with the same overall pass-rate are not equally usable. A `Rules & Facts` failure puts a wrong offer or a broken rule in front of a customer; a `Context` failure usually costs a repeated question; a `Tools` failure on a hang-up ends the call. GPT-6 Luna is the clearest case: 93.1 on the mandates, 70.0 on `Tools`. Qwen3.8 27B fails `Delivery` on 30% of turns. The three Indian open-weight models (Gnani Evon V3.3 30B A3B, Sarvam 105B, Sarvam 30B) pass `Tools` and `Rules & Facts` at 83 to 95% but lose `Goal` and `Context` on a third to a half of turns; Sarvam 30B passes `Goal` on only 46.6%. | |
| ### By slice | |
| <picture> | |
| <source media="(prefers-color-scheme: dark)" srcset="https://huggingface.co/datasets/Squadstack/conversational-llm-benchmark/resolve/main/assets/11-overall-by-slice-dark.png"> | |
| <img alt="Heatmap of overall pass-rate for 11 systems by campaign domain and target language; Tamil, Telugu, Kannada and Other are hatched as below the 40-call floor." src="https://huggingface.co/datasets/Squadstack/conversational-llm-benchmark/resolve/main/assets/11-overall-by-slice-light.png"> | |
| </picture> | |
| The ordering holds in every domain and every language with 40 or more calls. Marketplace is the hardest domain for the weaker systems (Gnani Evon 64.4, Sarvam 105B 63.2, GPT-4.1 62.9, Nemotron 60.1, Sarvam 30B 53.2). Tamil, Telugu and Kannada are below the 40-call floor and descriptive only. | |
| --- | |
| ## Uses | |
| ### Direct use | |
| Comparing LLMs for deployment as the reasoning core of an Indian outbound voice agent. The claim this benchmark licenses you to make is narrow: | |
| > *On next turns of Indian outbound voice-agent calls across loan sales, qualification, listing capture, counselling, service booking and survey campaigns, system A passed X% of rubric verdicts (95% CI …) and started its reply a median Y ms after the request, on the setup in LATENCY.md.* | |
| It also supports per-parameter diagnosis: which systems lose the flow, which invent facts, which reply in the wrong language, which mishandle tool calls. | |
| ### Out-of-scope use | |
| - **Training, fine-tuning or distilling any model.** This is a licence violation, not a recommendation: see [Licensing](#licensing). | |
| - **Re-identifying any caller, or recovering redacted text.** | |
| - **Claims about LLM quality in general**, in other languages or domains, about multi-turn behaviour, or about adversarial robustness. | |
| - **Claims about a slice under 40 calls**: Tamil, Telugu, Kannada and tool-only turns. | |
| - **Procurement decisions on latency alone.** Reproduce the measurement from your own network location before acting on it. | |
| --- | |
| ## What would make these numbers wrong | |
| | Risk | Status | | |
| |---|---| | |
| | Scores were computed before redaction | The prompts, histories and replies the systems and the judge saw contained the real names; the shipped files replace them with placeholders ([Redaction](#redaction)). Counting the shipped verdicts reproduces every number; a re-run on the shipped files sees slightly different input. Reported in [DATASHEET.md](docs/DATASHEET.md) | | |
| | The judge is an LLM | Claude Haiku 5.5, pinned, temperature 0, every verdict and reason stored. Differences between systems smaller than about one point should not be read as differences ([METHODOLOGY.md](docs/METHODOLOGY.md) §7) | | |
| | The judge shares a family with a candidate | No: no Anthropic model is on the board | | |
| | Turns without a verdict are dropped | Sarvam 105B lacks 439 of 9,962, Nemotron 348, Sarvam 30B 295, GPT-4.1 239, Gnani Evon 235, Gemma 4 26B 202, DeepSeek 3. If the dropped turns are harder than average, those scores are high. Reported per system in [LEADERBOARD.md](docs/LEADERBOARD.md) | | |
| | The history was produced by SquadStack's production agent | Every system answers from a history in that agent's style; disclosed, and part of why the in-house model is marked ★ | | |
| | The in-house model may have trained on these calls or campaigns | Checked at call and customer level before release; the campaigns' prompts and style are still those the model was tuned on, which is why it is marked ★ | | |
| | Campaign prompts were tuned for the in-house model | Disclosed; both variants were run and the Optimized variant helps every system ([PROMPTS.md](docs/PROMPTS.md)) | | |
| | Each turn is judged on the production history | Measures next-turn quality, not drift across a call | | |
| | One campaign per domain | Domain slices are campaign slices | | |
| | Small slices | 40-call floor; Tamil, Telugu and Kannada are below it | | |
| | Latency hosting differs | Maya V1 on a dedicated endpoint, OpenAI models on the OpenAI API, Gemma, DeepSeek, Qwen and Nemotron through OpenRouter providers, the Gnani and Sarvam models on Hugging Face Inference Endpoints; not comparable across systems ([LATENCY.md](docs/LATENCY.md)) | | |
| | Open-weight models ran quantized | fp8 for both Gemma 4 sizes, Qwen, DeepSeek, Gnani Evon and both Sarvam models, bf16 for Nemotron, as served by the provider ([`systems.json`](data/results/systems.json)) | | |
| --- | |
| ## Scope of improvements | |
| - **Transitioning from private to public prompts.** Move towards publicly accessible, standardized prompts and their corresponding conversations, with the goal of making the benchmark open, reproducible, and accessible to anyone looking to evaluate their models. | |
| - **Comprehensive evaluation framework.** Develop a broader, more rigorous evaluation rubric that goes beyond prompt adherence to assess sales effectiveness, human-like conversational behaviour, persuasion, adaptability, and overall interaction quality. | |
| - **Evaluating prompt dependency vs. model intelligence.** Understand how effectively the LLM can independently reason, adapt, and drive conversations towards desired outcomes without relying on explicitly scripted behaviours. | |
| --- | |
| ## Licensing | |
| Four separate statements, because they genuinely differ: | |
| | Component | Licence | | |
| |---|---| | |
| | **Conversation data** (histories and system replies) | Benchmark Evaluation Licence 1.0: evaluation only, no redistribution, no training, no re-identification. | | |
| | **Rubric, judge verdicts and campaign prompts** | CC BY-NC 4.0, attribution to SquadStack | | |
| | **Compilation, metadata and schema** | CC BY 4.0 | | |
| | **Evaluation code** (harness, judge and scoring, supplied on request) | Apache 2.0 | | |
| Full text in [`LICENSE.md`](LICENSE.md); the access terms you accept at the gate are in [`TERMS_OF_USE.md`](TERMS_OF_USE.md). Commercial evaluation is permitted; publishing results is encouraged; training is not permitted under any of the four. | |
| --- | |
| ## Citation | |
| ```bibtex | |
| @dataset{conversational_llm_benchmark_2026, | |
| title = {SquadStack Conversational LLM Benchmark for Voice Agents}, | |
| author = {{SquadStack MLOps Team}}, | |
| year = {2026}, | |
| version = {1.0.0}, | |
| publisher = {SquadStack}, | |
| url = {https://huggingface.co/datasets/Squadstack/conversational-llm-benchmark}, | |
| note = {Evaluation-only benchmark of 9,982 next-turn prompts from 471 Indian outbound | |
| voice-agent calls, scored by a 7-parameter LLM-judged rubric} | |
| } | |
| ``` | |
| SquadStack MLOps Team (2026). *SquadStack Conversational LLM Benchmark for Voice Agents* (Version 1.0.0) [Data set]. SquadStack. | |
| --- | |
| ## Acknowledgements | |
| The rubric draws on [G-Eval](https://arxiv.org/abs/2303.16634) for judge scoring and [MT-Bench / Chatbot Arena](https://huggingface.co/spaces/lmsys/mt-bench) for the single-output framing. The reproducibility controls for LLM-judged metrics borrow from [Sarvam's llm_wer](https://github.com/sarvamai/llm_wer). Card structure, figure system and comparability rules follow SquadStack's [Conversational Streaming ASR Benchmark](https://huggingface.co/datasets/Squadstack/conversational-streaming-asr-benchmark). | |
| **Dataset card authors** SquadStack MLOps Team | |