Dolly
Dolly is an open-source instruction-following LLM you can fine-tune on a single A100 in about 30 minutes.
Dolly is still one of the cheapest legal paths into instruction tuning: Apache 2.0 weights, a reusable 15k dataset, and a 30-minute fine-tune on one A100. But Databricks archived the GitHub organization on October 22, 2023, and the model itself is openly not state-of-the-art — weak on code, math, dates, and open-ended QA. Reach for it to learn or to ship a legally safe demo; pick a maintained open-weight model for anything you have to keep running.
Verified 3d ago · liveness 67/100 · cite: rightaichoice.com/tools/dolly
- Researchers learning how instruction tuning works on a single-GPU budget
- Developers needing a commercially licensed open-weight LLM for demos or prototypes
- Hobbyists training their first instruction-following model end to end
- Teams that need to audit the training data and pipeline for compliance
- Production applications that require ongoing maintenance or security patches
- Tasks that lean on code generation, math, dates, or multi-step reasoning
- Anyone who needs vendor support or an active community answering issues
We scan live Reddit threads, YouTube comments, X posts, G2 reviews and other communities — and hand you an honest verdict in under a minute.
- Honest verdict, not marketing
- Real pros & cons from real users
- Attributed quotes with receipts
3 free scans · no card needed
Skip Dolly if you need a model that is still maintained — Databricks archived the GitHub organization on October 22, 2023, so no fixes, security patches, or issue support are coming, and the model misses on code, math, and dates.
The weights are free, but you pay your own GPU bill: training the 12B model needs an A100, and inference on an A10 (g5.4xlarge) or V100 (p3.2xlarge) is billed by the hour by your cloud provider.
Dolly itself costs $0 — Apache 2.0 weights, public training code, and a CC-BY-SA dataset, so there is no subscription to compare. Your real spend is GPU time: an A100 hour for training, or an A10/V100 hour for the inference paths the README documents. That makes it cheaper than any hosted API for a one-off learning project, but more expensive in engineering time than an actively maintained open-weight model you can just pull and serve.
In short
Dolly — Dolly is an open-source instruction-following LLM you can fine-tune on a single A100 in about 30 minutes. Best for Researchers learning how instruction tuning works on a single-GPU budget, Developers needing a commercially licensed open-weight LLM for demos or prototypes, Hobbyists training their first instruction-following model end to end. Free to use.
What people actually say about Dolly — is it worth it?
We ran a structured research pass across product reviews, community discussions, and post-purchase forum threads to surface the patterns vendors won't publish themselves. Below: the recurring strengths, the hidden costs people mention most, and the cohort that consistently regrets adopting this tool.
8 mentions across 6 sources (Hacker News, YouTube, Product Hunt, Stack Overflow, GitHub, Lemmy), 66 more we could not attribute · researched Sep 9, 2026.
Weighted by the 74 posts each of 6 sources contributed.
- +Fine-tune 12B model in ~30 min on a single A100 GPU.
- +Open-source under Apache 2.0 with code and weights available.
- +Based on Pythia-12B with a clear architecture, good for learning.
- +Release of databricks-dolly-15k dataset under CC-BY-SA encourages experimentation.
- +Supports inference on A10 and V100 GPUs with quantization options.
- −Archived project, so no updates, support, or community feedback loops.
- −Struggles with programming, math, and complex reasoning tasks.
- −Lacks some abilities like well-formatted letter writing.
- −Setup can be tricky on non-standard hardware (e.g., Mac M1 issues).
- −Unresolved GitHub issues like torch.jit.trace support linger.
- • Hardware costs: A100 GPU for fine-tuning (~$2-3/hr cloud), A10/V100 for inference
- • Querying the model without proper setup may require technical expertise, time investment
Viability Score
How well maintained and how widely used is Dolly? Built from what the vendor actually publishes (docs, changelog, tutorials, integrations, pricing), whether the site is live, and how much real users discuss it. How we calculate this
Last calculated: October 2026
How we score →Key Features
- Instruction-following LLM derived from EleutherAI's Pythia-12B
- Fine-tuned on ~15k human-written instruction/response pairs (databricks-dolly-15k)
- Apache 2.0 license permits commercial use of the weights and training code
- Covers seven InstructGPT domains: brainstorming, classification, closed QA, generation, information extraction, open
- Trains the 12B model in roughly 30 minutes on a single A100 GPU
- Runs 12B inference on A10 GPUs (g5.4xlarge, 1x A10 24GB) with 8-bit weights via bitsandbytes
- Runs 12B inference on V100 GPUs (p3.2xlarge, 1x V100 16GB) in float16
- Loads through the Hugging Face transformers pipeline with trust_remote_code
- 12B weights published on Hugging Face as databricks/dolly-v2-12b
- Training code public on GitHub (databrickslabs/dolly) with train_dolly.py
- Instruction dataset databricks-dolly-15k released under CC-BY-SA
- 2.8B and 6.9B parameter variants also published
- Repository includes config, data, examples, test and training directories
- GitHub organization archived by an administrator on October 22, 2023 — no further updates
About Dolly
Dolly is Databricks' open-source instruction-following language model, built for developers and researchers who want to fine-tune an LLM without a data-center budget. It derives from EleutherAI's Pythia-12B and was fine-tuned on roughly 15,000 instruction/response records — the databricks-dolly-15k dataset, written by Databricks employees — covering the seven capability domains from the InstructGPT paper: brainstorming, classification, closed QA, generation, information extraction, open QA, and summarization. Weights ship on Hugging Face as databricks/dolly-v2-12b, and the training code lives on GitHub under Apache 2.0, so you can inspect, modify, and ship it commercially. The instruction dataset carries a CC-BY-SA license. Hardware demands stay modest: training the 12B model fits on a single A100 in roughly 30 minutes, and the README documents lighter inference paths — 12B on an A10 (g5.4xlarge, 1x A10 24GB) with 8-bit weights via bitsandbytes, or on a V100 (p3.2xlarge, 1x V100 16GB) in float16. The 2.8B and 6.9B variants run as-is on smaller A10 instances, and the whole thing loads through the standard Hugging Face transformers pipeline. Databricks is direct about the limits: dolly-v2-12b is not a state-of-the-art model, and it struggles with syntactically complex prompts, programming problems, math, dates and times, open-ended QA, hallucination, enumerating lists of a specific length, and stylistic mimicry. It also lost some capabilities present in the base model, such as well-formatted letter writing. The corpus carries the usual web-data baggage, and the employee-written instructions reflect a narrow demographic. Position it as a teaching artifact for instruction tuning rather than a production dependency — actively maintained open-weight models are the alternative once you need something you have to keep running.
Behind the Verdict
Dolly is the model people fine-tune first, and that reputation holds. Apache 2.0 weights, a 15k instruction corpus released under CC-BY-SA, and training that finishes in about half an hour on a single A100 — there are not many ways to learn instruction tuning that cheaply and legally. Pick it when the goal is understanding, not shipping. If you want to see how a base model becomes an instruction follower, Dolly's repo gives you config, data, examples, training and test directories and a train_dolly.py you can actually read. Pass on it for production. An administrator archived the Dolly GitHub organization on October 22, 2023, which means no updates, no fixes, and no issue support. That is a hard stop for anything with an uptime expectation. Capability is the other reason to look elsewhere. Databricks says plainly that dolly-v2-12b is not state-of-the-art, and lists the misfires: complex syntax, programming problems, math, dates and times, open-ended QA, hallucination, length-constrained lists, stylistic mimicry. It also dropped letter writing that the base Pythia-12B handled. In practice the hardware story is friendlier than the capability story. The 12B runs on an A10 with 8-bit weights, or a V100 in float16, and the 2.8B and 6.9B variants fit smaller A10 instances without tricks. A few lines of Python through the transformers pipeline get you generating. Watch the data caveats before you build anything customer-facing. The Pile carries content many users would find objectionable, the model reflects it, and the databricks-dolly-15k instructions reflect a demographic that is not representative of the global population. Closest alternative: reach for an actively maintained open-weight model — one with a live issue tracker and a team still shipping checkpoints — if
Researching Dolly? Get your full AI stack in 60 seconds.
Free, no signup — tell us your goal and get tools matched to your budget & existing stack.
Real-world workflow fit
Concrete scenarios for the personas Dolly actually fits — and what changes day-one when you adopt it.
Clone databrickslabs/dolly from GitHub, load the databricks-dolly-15k records from the data directory, and run train_dolly.py on a single A100 following the README, then evaluate the resulting checkpoint on the seven InstructGPT domains.
Outcome: A working fine-tuned 12B checkpoint in roughly 30 minutes of training time, plus a hands-on understanding of the instruction-tuning pipeline you can reuse on your own data.
Pull databricks/dolly-v2-12b from Hugging Face through the transformers pipeline, load it in 8-bit on a g5.4xlarge A10 instance, and wire it to brainstorming and closed-QA prompts over an internal document set.
Outcome: A demonstrable assistant with no per-token vendor cost and an Apache 2.0 license that clears legal review, running on a single mid-range GPU.
Use Cases
- Fine-tune a custom instruction-following model on domain-specific data for internal tools
- Experiment with LLM alignment techniques using a small, fast-to-train base model
- Deploy a chatbot for brainstorming or closed QA on a private dataset
- Teach yourself modern LLM fine-tuning with a fully open-source pipeline
- Create a lightweight assistant for classification or information extraction tasks
- Run inference on a single A10 GPU for low-cost proof-of-concept
Models Under the Hood
as of 2026-10-17
Limitations
- Databricks states dolly-v2-12b is not a state-of-the-art generative language model and is not designed to perform competitively with more modern architectures or models trained on larger corpora.
- It struggles with syntactically complex prompts, programming problems, mathematical operations, factual errors, dates and times, open-ended question answering, hallucination, enumerating lists of a specific length, stylistic mimicry, and having a sense of humor.
- Databricks also notes it lacks some capabilities present in the base model, such as well-formatted letter writing.
- The model reflects its training corpora — Pythia's pretraining data is largely web-collected and contains content many users would find objectionable — and the employee-written instructions reflect a narrow demographic.
- The GitHub organization was marked archived by an administrator on October 22, 2023, so no further updates, fixes, or support are coming.
as of 2026-10-08
Verification history
We have re-verified Dolly 10 times since . Each pass re-reads the vendor's own pages and re-checks every listed field against that evidence; passes where nothing had changed are marked as such.
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
Showing the 6 most recent of 10 verification passes.
Free to cite with attribution — this page re-verifies continuously.
12-month cost
Project the real annual outlay, including the implied monthly cost when only an annual tier is published.
Vendor list price only. Add-on usage, seat overages, and contract minimums are surfaced under Hidden costs & gotchas.
Plans compared
For each published Dolly tier: who it actually fits, and what it adds vs. the previous tier. Cross-reference the cost calculator above for projected annual outlay.
Open Source
$0
Ideal for
Researchers, hobbyists, and developers who want a commercially licensed open-weight LLM they can fine-tune on their own GPU without paying a vendor.
What this tier adds
Starting tier — $0 for Apache 2.0 weights, training code, and the CC-BY-SA dataset; the only cost is the GPU compute you supply.
Where the pricing makes sense
The company stage and team size where Dolly's pricing actually pencils out — and where peers do it cheaper.
Dolly itself costs $0 — Apache 2.0 weights, public training code, and a CC-BY-SA dataset, so there is no subscription to compare. Your real spend is GPU time: an A100 hour for training, or an A10/V100 hour for the inference paths the README documents. That makes it cheaper than any hosted API for a one-off learning project, but more expensive in engineering time than an actively maintained open-weight model you can just pull and serve.
Setup time & first value
How long it actually takes to get something useful out of Dolly — broken out by persona, not the marketing-page minute.
Researchers and developers already comfortable with Python and Hugging Face: 1-2 hours to clone the repo, install requirements, and get a first generation, then roughly 30 minutes of A100 time per fine-tune. Teams new to transformers pipelines: budget a day to provision the right GPU instance (A100 for training, A10 or V100 for 12B inference) and work through the README's quantization paths.
Switching to or from Dolly
How to bring data in from common predecessors and how to get it back out — written for the switcher, not the buyer.
- →From a hosted LLM API: export your prompt/response logs into instruction pairs, fine-tune dolly-v2-12b on them, and serve the checkpoint on your own A10 or V100.
- →From Pythia-12B: apply the released databricks-dolly-15k fine-tuning recipe and the train_dolly.py script to convert a base model into an instruction follower.
- →From another open-weight model: swap the model identifier in your transformers pipeline, but expect to re-tune prompts because Dolly's failure modes on code, math, and dates differ.
- ↗To an actively maintained open-weight model: point your transformers pipeline at the newer checkpoint and re-run your evaluation set, since Dolly has had no updates since October 2023.
- ↗To a hosted API: keep Dolly as the prototype and move production traffic to a managed endpoint once you need code, math, or multi-step reasoning.
- ↗To your own fine-tune of a newer base: reuse the databricks-dolly-15k corpus (CC-BY-SA) as instruction data on a current base model.
Integrations
Resources & Guides
- Resourcegithub.com
Dolly · Dolly
Helpful link from github.com
- Resourcegithub.com
README · Dolly
Helpful link from github.com
- Resourcegithub.com
CONTRIBUTING · Dolly
Helpful link from github.com
- Resourcehuggingface.co
Dolly V2 12b · Dolly
Helpful link from huggingface.co
- Resourcedatabricks.com
Hello Dolly Democratizing Magic Chatgpt Open Models · Dolly
Helpful link from databricks.com
Tutorials & Learning
YouTube returned 6 videos for “Dolly”, and we withheld 6: 6 could not be judged, because “Dolly” is a single word that other videos use for other things. We are showing none, because we could not prove any of them are about Dolly.
Official links
Tools that pair well with Dolly
Common stack mates teams adopt alongside Dolly, with the specific reason each pairing earns its keep.
Mistral
Mistral is the sovereign AI stack: frontier open-weight models you can self-host, fine-tune, or run in-region.
Zhipu AI
Zhipu AI's GLM-5.3 coding model and MaaS platform deliver open-source SOTA front-end programming, 1M context, and autonomous agents.
Baichuan 7B
Baichuan-7B is a free 7B bilingual Chinese-English base LLM to fine-tune and serve on your own hardware.
Featured Head-to-Head Comparisons
Dolly vs Surge Ai
Dolly is a free, lightweight open-source LLM you can fine-tune yourself in 30 minutes, perfect for experimentation or building custom bots. Surge AI, in contrast, is a premium enterprise platform that delivers expert human feedback for training frontier models—if you need rigorous RLHF or benchmarks like those used by Anthropic, Surge is the clear choice. Choose Dolly for low-cost tinkering and Surge for production-grade alignment.
Dolly vs Praktika
If you want to improve your spoken language fluency with real-time corrections, Praktika is the clear choice—it’s a polished mobile app with AI tutors. But if you need a lightweight, openly-licensed LLM to fine-tune for custom instruction tasks, Dolly is unbeatable for its price (free) and speed of training. These tools serve entirely different needs, so your pick depends on whether you’re a language learner or a developer.
Alternatives to Dolly
View allMistral
Mistral is the sovereign AI stack: frontier open-weight models you can self-host, fine-tune, or run in-region.
Zhipu AI
Zhipu AI's GLM-5.3 coding model and MaaS platform deliver open-source SOTA front-end programming, 1M context, and autonomous agents.
Baichuan 7B
Baichuan-7B is a free 7B bilingual Chinese-English base LLM to fine-tune and serve on your own hardware.
Frequently Asked Questions
Used Dolly? Help shape our editorial sentiment research.