tool-adapter-2tool

LoRA adapter for Qwen/Qwen3-0.6B that classifies each user turn in an IVF lead-qualification call into one of two tool calls:

• acknowledge() — greeting/salutation, agreement/proceed, confirmation of a value already present in the Current State, pure decline-of-offer, identity confirmations (ji haan / ji nahi).
• fallback() — everything else: answers to slot-elicitation questions (yes/no, soft-affirmations, values), decline + new information, agreement + new information, repairs/corrections, knowledge requests, handoff, ambiguous/uncertain/off-topic turns, greetings that carry actionable info.

There is intentionally no set_state() tool — slot-elicitation answers go to fallback().

Training data

• train.jsonl — 1614 rows (296 acknowledge() / 1318 fallback())
• valid.jsonl — 100 held-out rows (0 user-turn collisions with train)
• valid_flows.jsonl — 100 rows from real multi-turn call flows; pure acknowledge() → acknowledge(), otherwise fallback()

Source sheets: ~/Downloads/ACK().xlsx (dataset, from old data - fallbacks, from old data - all acks).

Base model

• Qwen/Qwen3-0.6B

LoRA config

lora_config_tool_adapter.yaml — rank 16, scale 20.0, dropout 0.05, 16 layers, batch 2 / grad accum 4, 400 iters, lr 1e-5, max_seq_length 768, mask_prompt true, seed 42.

Results

Eval set	Accuracy
valid.jsonl (100 rows)	81.0%
valid_flows.jsonl (100 rows)	95.0%

Known failure mode: acknowledge() is under-predicted relative to fallback() (class prior is ~82% fallback). Undersampling fallback, higher iteration counts, and exact-row teaching all regressed accuracy vs. this baseline.

Usage

from mlx_lm import load, generate

model, tok = load("Qwen/Qwen3-0.6B", adapter_path="./adapters/tool_adapter")

prompt = (
    "Given the conversation history and current state, choose the correct tool for the user turn.\n"
    "Current User Turn: mai 32 ka hu\n"
    "In the current User Turn, the correct tool call is"
)
out = generate(model, tok, prompt=prompt, max_tokens=16)
print(out)  # expect: fallback()