Poison-DPO / README.md
nbeerbower's picture
Upload README.md with huggingface_hub
7d49511 verified
|
Raw
History Blame Contribute Delete
1.87 kB
metadata
license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - hemlock
  - dpo
  - orpo
  - preference
  - code
pretty_name: Poison-DPO

Poison-DPO

Preference pairs for de-biasing Hemlock code models away from Python/JS habits ("poison") and toward idiomatic Hemlock.

Each row is a minimal pair:

  • chosen — clean, idiomatic Hemlock that runs under the interpreter.
  • rejected — the same program with one Python/JS-ism injected, verified to fail (parse/runtime error) or produce different output.

Because chosen and rejected differ by exactly one idiom, the preference signal (DPO/ORPO odds-ratio) isolates the specific behavior to suppress.

Coverage

Targets the standard-library / idiom errors a Python/JS-pretrained base reaches for:

Hemlock (chosen) Python/JS-ism (rejected)
map.delete(k) map.remove(k)
map.has(k) k in map
map.keys() Object.keys(map)
arr.push(x) arr.append(x)
print(x) println(x)
`${x}` (template) f"{x}"
"...json..." '...json...' (rune-literal trap)
arr.length len(arr)
s.contains(x) s.includes(x)
parse() / stringify() loads() / dumps()
for (x in arr) for (i in range(len(arr)))
null / fn(x){...} None / lambda x: ...

Construction

Every pair was generated and validated through the Hemlock interpreter: the chosen must exit 0, and the rejected must fail or diverge. Pairs are sourced from authored idiom templates (deep, targeted coverage) and from corruption of the real hemlang/hemlock-formulary-SFT programs (breadth over actual stdlib usage).

Schema

{"prompt": "...", "chosen": "<clean hemlock>", "rejected": "<poisoned hemlock>"}

Usable directly with ORPO/DPO/SimPO/CPO/IPO (needs chosen + rejected).