domofon commited on
Commit
9b36f26
·
verified ·
1 Parent(s): f0e1319

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +99 -0
README.md ADDED
@@ -0,0 +1,99 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ language:
4
+ - en
5
+ task_categories:
6
+ - text-generation
7
+ tags:
8
+ - agentic
9
+ - software-engineering
10
+ - tool-use
11
+ - code
12
+ - distillation
13
+ - gasai
14
+ - slm
15
+ size_categories:
16
+ - 1K<n<10K
17
+ source_datasets:
18
+ - togethercomputer/CoderForge-Preview
19
+ pretty_name: Gasai-Agent-CoderForge
20
+ configs:
21
+ - config_name: default
22
+ data_files: train.jsonl
23
+ ---
24
+
25
+ # Gasai-Agent-CoderForge
26
+
27
+ **5,205 verified-resolved agentic software-engineering trajectories, re-authored into the Gasai harness
28
+ format for pretraining small language models.** Derived from `togethercomputer/CoderForge-Preview`
29
+ (`reward==1`, Python). Each row is one complete tool-use trace serialized as a single Gasai control-token
30
+ sequence in the **`gasai`** field.
31
+
32
+ > Sister dataset to [`GasaiAI/Gasai-Agent-5k`](https://huggingface.co/datasets/GasaiAI/Gasai-Agent-5k)
33
+ > (from nvidia/Open-SWE-Traces) — **byte-identical format**, same pipeline.
34
+
35
+ ## What it is
36
+
37
+ | | |
38
+ |---|---|
39
+ | Traces | **5,205** |
40
+ | Unique problems (instances) | **2,704** (≤2 rollouts kept per instance) |
41
+ | Total tokens (StarCoder-2 tok) | **~312M** |
42
+ | Tokens / trace | median ~60K |
43
+ | File | `train.jsonl` (~1.0 GB) |
44
+ | Code language | **Python** |
45
+ | Natural language | **English** |
46
+
47
+ ### Format (Gasai harness)
48
+ ```
49
+ <|bos|><|system|>{system}<|tools|>[{read,bash,edit,write}]
50
+ <|user|>{structured task}
51
+ <|assistant|><think>{dense reasoning}</think><|tool_use|>{"name","input"}<|eos|>
52
+ <|tool_result|>{"tool_use_id":"tc_001","name","content"}<|eos|>
53
+ … (tool loop) …
54
+ <|assistant|><think>{reasoning}</think>{final answer}<|eos|>
55
+ ```
56
+ One `<think>` per assistant turn; `<|tool_use|>`↔`<|tool_result|>` paired by positional `tc_NNN`.
57
+ Schema: `{"trajectory_id": "...", "gasai": "<|bos|>…<|eos|>"}`.
58
+
59
+ ### Pipeline
60
+ 1. **Tool remap** — CoderForge's OpenHands tools (`execute_bash`, `str_replace_editor`, `finish`, `think`)
61
+ → the fixed Gasai 4-tool vocab `read` / `bash` / `edit` / `write`.
62
+ 2. **Constitution `<think>`** — the teacher's verbose reasoning re-written per turn into a dense, source-bound
63
+ `<think>` (every line a restated fact / grounded decision; no meta-narrative). Median ~42 tokens/think,
64
+ ~0.01% task-restatement.
65
+ 3. **Structured prompts** — first user turn restructured into a clean XML prompt (role/`<context>`, data in
66
+ tags, request last, embedded data in `{{ }}`). **80%** restructured; 20% keep the verbatim original
67
+ (fidelity fallback, never lossy).
68
+ 4. **Deterministic polish** — directory `read`s → `bash find`; dropped persistent-shell artifacts
69
+ (`C-c`/empty bash, hung "command is still running" steps); coerced object-typed edit/write args; stripped
70
+ emoji from finals (arrows kept); kept (incomplete) empty-final traces.
71
+ 5. **Dedup** — ≤2 rollouts per base instance (the source has up to 8 reward==1 rollouts per problem).
72
+
73
+ ## ⚠️ Known limitations (read before training)
74
+
75
+ This dataset was deep-audited (multi-agent semantic review). It is **pretrain-grade as one ingredient**, not a
76
+ clean SFT set. Honest caveats:
77
+
78
+ - **Benchmark contamination — NOT decontaminated.** Built from SWE-bench evaluation-harness images
79
+ (`sweb.eval.x86_64.*`); upstream mixes SWE-smith / SWE-rebench / R2E-Gym (the latter two are live
80
+ leaderboards). The 2,704 instances were **not** filtered against SWE-bench Verified/Lite or SWE-Gym test.
81
+ **Do not benchmark a model trained on this against those suites without first intersecting & removing
82
+ overlapping `instance_id`s** (it would be train-on-test).
83
+ - **`<think>` result-reading defect (~20–35% of traces, in-loss).** Some `<think>` blocks state "the previous
84
+ test failed" right after a *passing* run — the pytest `N passed` summary is buried in walls of `=` and the
85
+ distiller mis-read it. Affects the result-reading skill; fixable by regenerating the affected think blocks.
86
+ - **Oracle-test reliance (inherent).** Agents read the failing test, then implement exactly to it — valid
87
+ TDD, reward is honest, but the in-loss think leans on a test oracle that won't exist on real tasks, so the
88
+ *apparent* problem-solving capability is inflated.
89
+ - **Limited diversity.** ~312M tokens come from only **2,704 unique problems** (×≤2 rollouts); volume ≠
90
+ problem diversity. Mix with other sources for breadth.
91
+ - **Minor:** ~7% of traces edit test files (a "fix the test" anti-pattern); first step is uniformly
92
+ `bash find /testbed -maxdepth 2`; observations carry OpenHands wrappers (`[exit code]`, `cat -n` numbering)
93
+ and `/testbed` paths (conditioning only — tool observations are not in loss).
94
+
95
+ ## Source & license
96
+ Derived from [`togethercomputer/CoderForge-Preview`](https://huggingface.co/datasets/togethercomputer/CoderForge-Preview),
97
+ config `filtered_reward1` (the curated `reward==1` union of SWE-smith / SWE-rebench / R2E-Gym tasks).
98
+ Underlying repositories carry their own permissive licenses (MIT / Apache-2.0 / BSD; see the per-row
99
+ `license` in the source). Please cite CoderForge-Preview and the upstream task sets.