mondk/UltraData-SFT-Agent-2609-CLEANED

1. Summary

Metric	Value
Input records	21,006
Retained	20,850 (99.3%)
Discarded	156 (0.7%)
Truncated records	4,611 (22.1%)
Full sequence tokens (incl. template + tool definitions)	p50 = 676 · p90 = 1,902 · p99 = 1,990 · max = 2,000
JSON errors	0
Remaining Hanzi (easy tasks)	13 records

Format:

{
  "messages": [
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "", "reasoning_content": "...",
     "tool_calls": [{"type": "function",
                     "function": {"name": "web_search", "arguments": {"queries": ["..."]}}}]},
    {"role": "tool", "content": "..."}
  ],
  "tools": [
    {"type": "function",
     "function": {"name": "web_search",
                  "parameters": {"type": "object",
                                 "properties": {"queries": {"type": "array", "items": {"type": "string"}}},
                                 "required": ["queries"]}}}
  ],
  "domain": "Search_Agent-plan"
}

---

2. system + tools dominate the budget — the biggest issue

Measured by rendering each record through the real chat_template.jinja of MiniCPM5-1B, not by estimating chars / 4:

Source	system	tools block	Total if kept as-is
General_Agent	8,743	16,980 (37 tools)	p50 ≈ 50,473
Tool_Use	1,527	3,250 (15 tools)	p50 ≈ 5,990
Search_Agent	23	439 (3 tools)	p50 ≈ 903

For 2 of the 3 files the fixed overhead alone exceeded 2,000 tokens before a single character of conversation was counted. A "< 2,000 tokens per line" rule is therefore impossible unless system and tools are also addressed. Resolved by dropping the system message and optimizing the tools block.

2.1 Tool definitions: 57–97% of the fixed cost

Keeping only the functions actually called in the trace, with description, title, $schema, default and examples stripped at every schema level:

Source	tools: original → optimized	Reduction	Tools actually called
General_Agent	16,980 → 539	97%	4 / 37
Tool_Use	3,250 → 319	90%	3 / 15
Search_Agent	439 → 191	57%	1 / 3

Any tool_call naming a function with no definition causes the whole record to be discarded (tool_undef) — it cannot teach correct function calling.

2.2 The give-up filter was wrong by a factor of 180

The first version used a broad regex (i cannot, i can't, unable to, i'm not able to…) and flagged 360 records — all of them false positives. In agent traces these phrases occur constantly in valid turns:

• "No LibreOffice on this host, so I can't render a PDF preview, but the document is built." → successful trace.
• "I cannot process both a return and an exchange on the same order." → correct policy for a retail agent.
• "If you're unable to provide that, I can transfer you to a human agent." → correct business logic.

Narrowed to three simultaneous conditions — final message only, ≤ 90 tokens, and strong refusal phrases only (i (am) (unable|cannot) to (complete|fulfill|finish|proceed|continue|answer), i apologize, but i cannot, as an ai model, i don't have the ability to). Result: 360 → 2, both genuinely abandoned tasks.

2.3 Dropping the exchange next to the final answer creates contradictions

Unit-level dropping (an assistant plus its following tool responses) can remove the exchange immediately preceding the final answer, leaving an older, failed step as the last visible tool result. The trace then reads as nonsense: not found → "Everything checks out, here's what I built."

Protecting the second-to-last unit leaves 11 records (0.05%) still showing this pattern. The cost is that traces which cannot be compressed while keeping the tool chain intact are now discarded rather than degraded into a plain Q→A pair — which is why General_Agent retains 885 of its 1,006 records. In exchange, 883 of 885 (99.8%) General_Agent records keep at least one tool_call, and that share is essentially unchanged under either cap setting (§5).

---

3. Dataset composition

Property	Value
Records ending with a text answer	8,759 (42.0%)
Records ending with a tool_call only (no final answer)	12,091 (58.0%)
Records with a truncated task specification	242 (1.2%)
Exact duplicates	0
Repeated prompts (identical first user message)	327 groups — 0 true duplicates
<tool_response> orphaned at the start	0

58.0% end on a tool_call rather than an answer, and 56.7% (11,824) are single-step traces (user → assistant(reasoning + tool_call), exactly 2 messages), by design of the upstream corpus. They teach emitting a function call; they do not teach consuming a tool result to produce an answer. Repeated prompts are not duplicates: two traces can share a user question while differing in the tool-call signature (only 2 groups share an identical tool chain).

---

4. Filtering criteria applied

#	Criterion	Threshold / rule	Count
1	Malformed record	JSON parse, invalid role, no user/assistant, empty assistant without tool_calls, does not end on assistant	19
2	Abandoned task (give-up)	strong refusal phrase in the final message, ≤ 90 tokens	2
3	Degenerative loop	20-word block repeated ≥ 6 times	15
4	Cannot fit the budget while keeping the tool chain	any reduction order still > 2,000 tokens	86
5	Fixed overhead alone exceeds the cap	optimized tools block > 2,000 tokens	34
6	tool_call to an undefined function	name absent from the tools list	0
7	Exact duplicate	md5 over the full messages list	0

---

5. Truncation

• tool_calls arguments are never truncated — they are the learning target. Truncating them would teach the model to emit broken function calls.
• Messages are never dropped individually — that would leave orphaned <tool_response> blocks. Whole units are dropped.
• Reduction order, least harmful first: tool response (input, ≤ 150 tok) → user prompt (input, ≤ 700 tok) → reasoning (target, ≤ 600 tok) → assistant answer (target, ≤ 800 tok) → drop middle units (heaviest first) → trim head/tail.
• Cuts land on paragraph / sentence boundaries, never mid-sentence.
• Cuts auto-balance ``` fences: if the cut falls inside a code block, back off to before the opening fence, or close it. Result: 0 cuts inside a code block, 1 unbalanced fence (pre-existing in the source data).
• The first user message, the final assistant message, and the unit immediately preceding it are always preserved.

22.1% of records carry a cut mark in their final content (22.6%, i.e. 4,719 records, if cuts inside reasoning_content are counted as well), overwhelmingly in tool responses (inputs) capped at 150 tokens to make room for more agent steps. Raising the tool-response cap from 150 → 300 lowers that broader truncation rate to 14.7% (4,719 → 3,057 records, 20,834 retained), but also lowers the mean message count (6.6 → 6.1). The share of General_Agent records retaining a tool_call stays at 99.8% under both settings.

All figures in this report were re-derived from the output file (stats_agent.py, verify_report.py, exp_cap300.py, audit_agent.py, audit_wrong.py), not copied from build logs.