thenukegun10x/LFM2.5-24B-A2B-BF16-Reasoning-Traces

LFM2.5-24B-A2B BF16 Reasoning Traces

1,242 accepted teacher reasoning conversations: 605 GSM8K math examples and 637 Glaive single-tool examples. Every accepted target has a separate reasoning trace. This preserves the original mixed reasoning batch; it is not an all-math collection. The companion parallel-tool dataset contains a separate 1,200-example collection.

The original candidate bank contained 10,000 tasks (5,000 from each source). A bounded collection produced these 1,242 accepted records; not all bank tasks were attempted. GSM8K prompts came only from the original training split. Original test questions were used only for exclusion screening and are not published here. Glaive examples use the first labelled single tool call and preserve any earlier clarification turns. Prior history is not a newly generated reasoning target.

The collection was designed for short-context syntax/reasoning recovery. Its topic coverage is limited to grade-school math and synthetic single-tool selection, not broad general reasoning. There are 392,210 scored tokens, with no truncation and a maximum permitted context of 2,048 tokens.

Subset	Train	Validation	Total
math	535	70	605
single_tool	574	63	637
all	1,109	133	1,242

Generation and verification

The teacher was LiquidAI/LFM2.5-24B-A2B, pinned to revision 1769516ac588c4c558089c94f27d96b0dc67b183, running in BF16 without weight quantization on a Modal NVIDIA H100 80GB worker. The collection runtime used vLLM 0.21.0 and PyTorch 2.11.0+cu130. These are generated teacher continuations, not human-written explanations.

The native generation prompt supplied the opening <think> tag. Accepted continuations supplied one closing </think> tag and a complete final answer or native tool-call block. The exported text includes the complete prompt and both thinking boundaries. Generation was not forced to emit reference-answer text and did not use a constrained output grammar. Ground-truth labels were used for rejection checks, not included in the user prompt.

Automatic checks validate the final numeric answer or exact typed tool names and arguments, native syntax, tool schemas, explicit calculation annotations when present, completion boundaries, and context length. They do not prove every sentence of reasoning or every upstream source label. Ordinary arithmetic written outside explicit calculation annotations is not exhaustively verified. Tools were not executed; there are no invented tool results. This is a training/distillation collection, not a new independent benchmark.

Format and use

from datasets import load_dataset
import json

ds = load_dataset("thenukegun10x/LFM2.5-24B-A2B-BF16-Reasoning-Traces", "all")
row = ds["train"][0]
tools = json.loads(row["tools"])
messages = row["messages"]
for message in messages:
    for call in message["tool_calls"]:
        call["function"]["arguments"] = json.loads(call["function"]["arguments"])

reasoning and answer expose the final assistant turn directly; messages preserves the complete conversation, including any earlier clarification turns. In Parquet, tools, source_metadata, and provenance are JSON strings, and each function's arguments is a JSON string. This keeps heterogeneous API schemas and argument types compatible with Arrow and the dataset viewer. Decode those fields with json.loads for structured use. No argument types or values are intentionally changed.

text is the original model-native rendered training text, with <think>...</think> and, for tool examples, <|tool_call_start|>[function(...), ...]<|tool_call_end|>. It is not a generic model's chat template. When adapting these traces to another model, use that model's template and preserve the distinction between reasoning and the final answer/calls.

The public validation split is the existing local development split, renamed without moving any rows. It is not the original source benchmark test split. Source groups do not cross the released train/validation boundary.

The following additional downloads preserve the frozen preparation artifacts:

• native/conversations.jsonl: original structured records, including dictionary arguments and full tool definitions.
• native/examples.jsonl: original rendered examples, token IDs, loss masks, and per-token parts, in original row order.
• tokenized/recovery2048.npz: ids (int32), mask (bool), and split (uint8; 0=train, 1=validation), padded to 2048 tokens. Rows align with native/examples.jsonl, not the subset-sorted Parquet files. False mask positions exclude prompts/history/padding from scored loss.
• native/chat_template.jinja: teacher-native formatting reference, accompanied by the teacher license.
• provenance/original_preparation_manifest.json, provenance/release_audit.json, and SHA256SUMS: source integrity and release checks.

These are token/continuation datasets. No teacher logits or probability cache are included, and this release did not start QAD training. The archived preparation manifest's pending-teacher-targets stage refers to a possible later logits/distillation pass, not missing conversation records in this release.

Attribution and terms

GSM8K source material is MIT-licensed by OpenAI; Glaive source material is Apache-2.0-licensed by Glaive AI. The repository uses component licensing, with the source license attached to each record. Newly contributed packaging and generated continuations are offered under Apache-2.0 to the extent the publisher holds rights in them, subject to the retained upstream terms.

Liquid AI is the teacher model author. The teacher model and the copied chat template have their own LFM Open License; this release does not distribute model weights. Dataset source licenses are recorded per row and are not replaced by the teacher license. See LICENSE.md and NOTICE.md for the component terms and modification notices. This is a community release by thenukegun10x, without endorsement by the source dataset or model authors.