Skip to content

Automation foundation model for tiny devices: 2-bit, 8-29 MB, tool calls, ASR, structured extraction and embeddings on phones, wearables, smart homes, robots, cars and microcontrollers.

fix tool calls

HenryNdubuakue36b1f181ffe
diff --git a/scripts/tools_data.py b/scripts/tools_data.py
index 53ccd4c..8308b9b 100755
--- a/scripts/tools_data.py
+++ b/scripts/tools_data.py
@@ -22,7 +22,7 @@ Datasets and their native formats:
3. glaiveai/glaive-function-calling-v2
Tool definitions embedded in the system prompt; chat is a flat string with
USER:/A: role markers and <functioncall> tags. Examples with empty answers
- (answers=[]) are filtered out in the final combine step.
+ (answers=[]) are kept so the model learns when NOT to call functions.
4. Team-ACE/ToolACE
Tool definitions as a JSON array in the system prompt; function calls use a
@@ -227,6 +227,26 @@ def _parse_toolace_calls(text):
return answers
+def compact_tools(tools_json):
+ """Strip verbose schema descriptions, keep only function name + param name/type."""
+ try:
+ tools = json.loads(tools_json)
+ except (json.JSONDecodeError, TypeError):
+ return tools_json
+ compact = []
+ for t in tools:
+ params = {}
+ raw_params = t.get("parameters", {})
+ props = raw_params.get("properties", raw_params) if isinstance(raw_params, dict) else {}
+ for pname, pdef in props.items():
+ if isinstance(pdef, dict):
+ params[pname] = pdef.get("type", "string")
+ else:
+ params[pname] = str(pdef)
+ compact.append({"name": t["name"], "parameters": params})
+ return json.dumps(compact)
+
+
CONVERTERS = {
"xlam-function-calling-60k": convert_xlam,
"APIGen-MT-5k": convert_apigen_mt,
@@ -252,15 +272,17 @@ def load_and_combine():
parts.append(converted)
combined = concatenate_datasets(parts)
- pre_filter = len(combined)
- # Filter out examples with empty answers
- combined = combined.filter(
- lambda ex: json.loads(ex["answers"]) != [],
- desc="Filtering empty answers",
+ # Compact tool schemas: strip descriptions, keep only name + param name/type
+ print(f"\nCompacting tool schemas...")
+ combined = combined.map(
+ lambda ex: {"tools": compact_tools(ex["tools"])},
+ desc="Compacting tools",
)
+
+ empty_count = sum(1 for ex in combined if json.loads(ex["answers"]) == [])
print(f"\n{'='*60}")
- print(f"Combined dataset: {len(combined)} examples (filtered {pre_filter - len(combined)} empty answers)")
+ print(f"Combined dataset: {len(combined)} examples ({empty_count} with empty answers = no-call examples)")
print(f"Columns: {combined.column_names}")
counts = Counter(combined["source"])
diff --git a/src/data.py b/src/data.py
index 319dbcc..4ebc294 100644
--- a/src/data.py
+++ b/src/data.py
@@ -173,27 +173,6 @@ def _cache_key(prefix, n_samples, max_enc_len, max_dec_len):
return hashlib.md5(key.encode()).hexdigest()[:12]
-def _compact_tools(tools_json):
- """Strip verbose schema descriptions, keep only function name + param name/type."""
- import json
- try:
- tools = json.loads(tools_json)
- except (json.JSONDecodeError, TypeError):
- return tools_json
- compact = []
- for t in tools:
- params = {}
- raw_params = t.get("parameters", {})
- props = raw_params.get("properties", raw_params) if isinstance(raw_params, dict) else {}
- for pname, pdef in props.items():
- if isinstance(pdef, dict):
- params[pname] = pdef.get("type", "string")
- else:
- params[pname] = str(pdef)
- compact.append({"name": t["name"], "parameters": params})
- return json.dumps(compact)
-
-
def prepare_tool_call_pairs(ds, tokenizer, max_enc_len=256, max_dec_len=1024):
"""Prepare tool-call encoder-decoder pairs with <tool_call> task token.
@@ -221,9 +200,9 @@ def prepare_tool_call_pairs(ds, tokenizer, max_enc_len=256, max_dec_len=1024):
eos_id = tokenizer.eos_token_id
tool_call_id = tokenizer.tool_call_token_id
- # Encoder: query only. Decoder prefix: compact tools. Decoder answer: answers.
+ # Encoder: query only.
enc_texts = [ex["query"] for ex in ds]
- tools_texts = [_compact_tools(ex["tools"]) for ex in ds]
+ tools_texts = [ex["tools"] for ex in ds]
ans_texts = [ex["answers"] for ex in ds]
num_workers = min(os.cpu_count() or 1, 8)
diff --git a/src/run.py b/src/run.py
index 87b2b1d..0dba654 100644
--- a/src/run.py
+++ b/src/run.py
@@ -30,8 +30,6 @@ def generate(model, params, tokenizer, query, tools="[]", max_gen_len=512, seed=
Encoder: query only.
Decoder: prefilled with [BOS, <tool_call>, tools_tokens...], then greedy decode.
"""
- from .data import _compact_tools
-
enc_tokens = tokenizer.encode(query)
enc_input = jnp.array([enc_tokens])
@@ -44,8 +42,8 @@ def generate(model, params, tokenizer, query, tools="[]", max_gen_len=512, seed=
{"params": params}, enc_input, src_mask=src_mask, method="encode"
)
- # Build decoder prefix: [BOS, <tool_call>, compact_tools_tokens...]
- tools_tokens = tokenizer.encode(_compact_tools(tools))
+ # Build decoder prefix: [BOS, <tool_call>, tools_tokens...]
+ tools_tokens = tokenizer.encode(tools)
prefix = [eos_id, tool_call_id] + tools_tokens
prefix_len = len(prefix)