fix(halo): lower deepseek sampling temp to 0.7 to break reasoning loops #6

Merged

harald merged 1 commit from hermes/loop-guard-temp into main

2026-08-08 13:40:02 +02:00

AGit

hermes commented

2026-08-08 13:32:59 +02:00

Contributor

repeat-penalty was dropped earlier because it crippled tool calls (it dampens
repeated tokens indiscriminately, including JSON tool bodies). Replace it with
lower temp 0.7 / top-p 0.95, which removes the sampling wander in free-text
generation where the loops actually occur, without touching tool-call bodies.
n-predict stays as the deterministic backstop.

repeat-penalty was dropped earlier because it crippled tool calls (it dampens repeated tokens indiscriminately, including JSON tool bodies). Replace it with lower temp 0.7 / top-p 0.95, which removes the sampling wander in free-text generation where the loops actually occur, without touching tool-call bodies. n-predict stays as the deterministic backstop.

hermes added 1 commit

2026-08-08 13:32:59 +02:00

fix(halo): lower deepseek sampling temp to 0.7 to break reasoning loops 7428fce0a0

repeat-penalty was dropped earlier because it crippled tool calls (it dampens
repeated tokens indiscriminately, including JSON tool bodies). Replace it with
lower temp 0.7 / top-p 0.95, which removes the sampling wander in free-text
generation where the loops actually occur, without touching tool-call bodies.
n-predict stays as the deterministic backstop.

hermes force-pushed hermes/loop-guard-temp from 7428fce0a0 to 3b879d8888

2026-08-08 13:39:14 +02:00

Compare

harald approved these changes

2026-08-08 13:39:58 +02:00

harald merged commit 1873e5de53 into main

2026-08-08 13:40:02 +02:00

harald referenced this pull request from a commit

2026-08-08 13:40:03 +02:00

Merge pull request 'fix(halo): lower deepseek sampling temp to 0.7 to break reasoning loops' (#6) from hermes/loop-guard-temp into main