speech-llm-eval-dataset

Evaluation dataset for a Hindi-speaking voice collections agent ("वाणी", Bajaj Finserv EMI-bounce calls). Built from a stratified sample of 1000 conversation turns (out of 6351 total) in bajaj-convo-hindi.json.

Each record is a prompt only — the final agent response is intentionally excluded; responses are generated separately at evaluation time.

Files

• prompt_hindi.json — source scenario templates (system prompt + per-scenario injection prompts, with {placeholder} fields) used to build every record's system prompt.
• prompt_hindi_text.json — 1000 records, plain text. Each record has a messages list (system + prior conversation history + final user turn as text), plus context, scenario, guidelines, and id.
• prompt_hindi_audio.json — same 1000 records, except the final user message's content is an OpenAI-multimodal-compatible input_audio block (base64-encoded WAV) instead of text. Prior history turns remain text.
• audios/<id>.wav — the synthesized speech for each record's final user turn (mono, 16-bit PCM), generated via a Bajaj-internal FlowTTS endpoint (voice: tara). Referenced by audio_path in prompt_hindi_audio.json.
• sampled_records.json — the raw 1000 sampled records (unmodified), for provenance.

Record schema (prompt_hindi_text.json / prompt_hindi_audio.json)

{
  "id": "dr-...",
  "conversation_id": "...",
  "persona": "green",
  "scenario": "FTB | general | STAB | FEMI",
  "context": { "customer_name": "...", "overdue_amount": "...", "...": "..." },
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "assistant", "content": "..."},
    {"role": "user", "content": "..."},
    "...",
    {"role": "user", "content": "..." } // text version
    // OR, in prompt_hindi_audio.json:
    {"role": "user", "content": [{"type": "input_audio", "input_audio": {"data": "<base64 wav>", "format": "wav"}}]}
  ],
  "user_query": "plain text of the final user turn (kept in both files for reference)",
  "guidelines": ["...evaluation criteria for grading the agent's next response..."],
  "audio_path": "audios/<id>.wav"  // audio file only
}

Sampling

1000 records were sampled from the 6351-record source with proportional stratification by scenario (seed 42), so the scenario mix matches the full dataset: FTB ~91%, general ~7.5%, STAB ~1.3%, FEMI ~0.3%.

Build pipeline

See scripts/:

1. 01_sample.py — stratified sample → sampled_records.json
2. 02_build_text_prompts.py — fill templates → prompt_hindi_text.json
3. 03_generate_audio.py — TTS the final user turn → audios/*.wav
4. 04_build_audio_prompts.py — embed base64 audio → prompt_hindi_audio.json
5. 05_push_to_hf.py — push this repo to the Hugging Face Hub