tyfeng1997/gemma4-e4b-swe-sft-32k

Gemma E4B SWE SFT — response-level, competition 32K context

1,737 train examples and 159 val examples, derived from 140 selected successful DeepSeek Flash max rollouts in tyfeng1997/gemma4-swe-deepseek-flash-max-verified-trials. Train contains 131 rollouts on 43 tasks; val contains 9 rollouts on 3 tasks. Splitting is deterministic by task ID; trajectories of the same task do not cross train and val.

Source / 来源

Original SWE tasks come from the public/development task dataset supplied with Google — The Gemma 4 Developer Agent Competition on Kaggle. Teacher completions were generated by deepseek-flash, thinking enabled, reasoning effort max, using the original default tool suite, analyzer agent and unchanged system prompts. The parent dataset retains the full original trials and their reward-verification evidence. These are not hidden/private evaluation tasks. Upstream source terms apply; see TERMS.md.

Transformation / 处理方式

Each assistant response creates one example. The prefix retains system/task messages, historical assistant visible text and tool calls, and tool observations. Historical thinking is explicitly removed (reasoning, reasoning_content, thinking_blocks). The current target retains its reasoning, visible text and tool calls when present. Labels for the entire historical prefix are -100; only the current response, including required template control tokens, contributes to the loss.

Input and target are encoded using the official tokenizer and chat template from google/gemma-4-E4B-it-qat-w4a16-ct at revision 6cd26aaa2357fb2bad8c51699a7558a4d1a965bb. The maximum is 32,768 tokens for the full input + current target, including system instructions, tool definitions, observations and template control tokens. This is the competition serving/evaluation configuration's context limit, NOT a statement of Gemma E4B's native maximum context length. Oversized responses are excluded without hard truncation; 303 such responses were excluded across all positive trials, including the review group. The longest train example has 32,741 tokens; the longest val example has 32,612 tokens.

The raw teacher generated its continuation while receiving historical thinking. Removing that history for the student is a distillation transformation, not a claim of identical teacher/student contexts. Student inference should use the same historical-thinking removal policy while allowing current-step thinking.

Fields and loading

input_ids and labels are ready for causal-LM SFT with the matching tokenizer. Preserve the supplied labels and use a collator that pads labels with -100; a generic collator that regenerates labels from input IDs would incorrectly train on the historical prefix. sequence_length and supervised_tokens describe the exact encoded example. source_trial_id joins to the parent dataset.

prompt_messages_json, target_message_json, and tools_json are JSON strings to preserve heterogeneous tool argument schemas without Arrow type conflicts; decode them with json.loads for inspection or re-serialization. They contain the exact processed messages, with historical thinking removed and current target thinking preserved. Main and analyzer calls can both be represented.

import json
from datasets import load_dataset

ds = load_dataset("tyfeng1997/gemma4-e4b-swe-sft-32k", token=True)
train, val = ds["train"], ds["val"]
row = train[0]
prefix = json.loads(row["prompt_messages_json"])
target = json.loads(row["target_message_json"])
assert len(row["input_ids"]) == len(row["labels"]) <= 32768

Quality limits

Only the existing auto-screened or manually accepted candidate group is included. The 2,752 response examples from 125 review/rejected trajectories are not part of train or val. Passing the reward verifier and automated screening does not establish a full manual correctness audit. The dataset may still contain unhelpful actions, imperfect reasoning or fixes with untested behavior. No student model has been trained as part of this release.